uni.data

uni.Mat is a high-performance, NumPy-compatible matrix library for Scala 3.7.0+.

It provides a zero-overhead, type-safe interface for scientific computing by leveraging Scala 3 Opaque Types. Designed for developers who need the ergonomics and reproducibility of the Python/NumPy ecosystem within the JVM, uni.Mat features 100% faithful implementations of NumPy's random generation and strided array logic.

Key Features

Zero-Overhead Types: Uses opaque type Mat[T] = Internal.MatData[T] to ensure that at runtime, your matrices are as lean as raw arrays, with no wrapper object overhead.
NumPy Random Fidelity: 1:1 behavioral matching for rand, randn, uniform, randint, etc. using a high-performance PCG64 implementation.
Strided Memory Layout: Supports rowStride and colStride, enabling $O(1)$ transpose and zero-copy slicing/views, mirroring NumPy's internal engine.
Broadcasting & In-place Ops: Built-in support for NumPy-style broadcasting and memory-efficient in-place mutation operators (:+=, :-=, :*=, :/=).
Deep Learning Primitives: Activation functions (sigmoid, relu, softmax, leakyRelu) use parallel fork/join for contiguous matrices — outperforming NumPy's single-core SIMD on multi-core hardware.
Pandas-Style Data Analysis: head/tail, shift/pct_change, rolling, describe, fillna, idxmin/idxmax, cummax/cummin, nlargest/nsmallest, between, valueCounts, and named-column CSV access via MatResult.

For high-accuracy scientific modeling or other applications requiring extreme precision, uni.Mat provides a Big numeric type.

High Precision: Big Type Guide — Learn about high-precision matrices, and how to use Mat[Big].

Rust companion crate (`rust/`)

The rust/ directory holds a Rust port of the numerically heavy parts of the library. It is a standalone crate (t3prf), not a JNI binding — nothing in the Scala library calls into it. It exists so the same algorithms can be run outside the JVM and so the Scala implementations have an independent second opinion to be checked against.

Bit-identical NumPy random numbers. NumPyRng reproduces np.random.default_rng(seed) — PCG64 XSL RR 128 behind NumPy's SeedSequence expansion — and therefore also reproduces uni.data.NumPyRNG draw for draw. Seed it with 0 in Scala, Rust or Python and all three yield the same stream. next_u64, next_i32, next_f64, uniform and next_bounded_u32 are exact by construction: integer arithmetic plus one exactly-representable division by 2⁵³, no tolerance involved.
randn too, via the same 256-layer Ziggurat, with the tables generated from the Scala source rather than transcribed. It matches NumPy bit-for-bit. Against the JVM it differs on roughly 2.5 draws per million — always a tail draw, always by exactly one ulp, because Math.log1p and C's log1p round a few arguments differently. Never a control-flow divergence, so the streams stay aligned rather than drifting apart.
3PRF (estimate_3prf_is_full, _oos_rec, _oos_cv) — the same three procedures as uni.stats.Tprf3, including the per-window cross-product downdates.

Both halves are pinned by cross-language parity fixtures that neither side can quietly move: test-data/numpy-rng-parity/ for the generator and test-data/tprf3-parity/ for 3PRF. Each is checked independently by a Scala suite and a Rust test against the same committed reference, so neither test needs the other language installed. The RNG fixtures additionally pin specific draws to NumPy's own bit patterns as absolute values — a check that regenerating the fixtures cannot launder.

cd rust && make all          # test, fmt, clippy, file-size check
cargo run --release --bin bench_tprf3

Visualization (`uni.plot`)

import uni.plot.* adds .scatter(), .hist(), and .plot() directly on MatD. Each method opens an interactive Swing window, or saves a PNG when saveTo is supplied. Pass a PlotStyle to control dimensions, colours, and export consistency.

#!/usr/bin/env -S scala-cli shebang -Wunused:imports -Wunused:locals -deprecation
//> using dep org.vastblue:uni_3:0.14.2
import uni.data.*
import uni.plot.*

val iris = MatD.readCsv("datasets/iris.csv")
// columns: sepal_length(0), sepal_width(1), petal_length(2), petal_width(3)

iris.scatter(2, 3, title = "Iris: petal length vs petal width")
iris.hist(bins = 20, title = "Iris: sepal length distribution")
iris.plot(title = "Iris: all 4 features",
          labels = Seq("sepal length", "sepal width", "petal length", "petal width"))

// PlotStyle — override only what you need; everything else defers to the GGPlot2 theme
iris.scatter(2, 3, style = PlotStyle(width = 1200, height = 800))
iris.hist(bins = 20, style = PlotStyle(background = Some(java.awt.Color.BLACK),
                                       foreground = Some(java.awt.Color.WHITE)))

// PlotStyle.uniform — identical 800×500 for all saved images
iris.scatter(2, 3, saveTo = "docs/images/iris-scatter", style = PlotStyle.uniform)
iris.hist(bins = 20, saveTo = "docs/images/iris-hist",  style = PlotStyle.uniform)

Petal length vs petal width Iris clusters

Sepal length distribution across 150 samples

See jsrc/iris.sc and jsrc/anscombe.sc for runnable demos, and Plot Guide for the full PlotStyle API.

Performance

Core matrix operations

NumPy: Python 3.14.6 / NumPy 2.4.6 (see py/bench.py). Breeze/MatD: uni 0.14.1 / Scala 3.8.2 / JVM 21 (see jsrc/bench.sc, jsrc/benchBreeze.sc); min times. Note: on this machine netlib's JNIBLAS could not load libopenblas.dll, so MatD and Breeze matmul ran on the pure-JVM fallback (Java11BLAS / VectorBLAS) while NumPy used native OpenBLAS — the matmul row compares fallback vs native. Where JNIBLAS loads (Linux and macOS tables below), matmul is level.

Operation	NumPy	Breeze	MatD
`randn(1000×1000)`	8.5 ms	21.6 ms	6.8 ms
`matmul 512×512`	0.79 ms	5.3 ms	3.2 ms
`sigmoid(1000×1000)`	9.0 ms	4.2 ms	0.73 ms
`relu(1000×1000)`	1.3 ms	2.9 ms	0.51 ms
`add(1000×1000)`	1.7 ms	1.2 ms	0.58 ms
`sum(1000×1000)`	0.13 ms	0.37 ms	0.03 ms
`mean(1000×1000)`	0.14 ms	3.6 ms	0.03 ms
`std(1000×1000)`	2.4 ms	5.4 ms	0.43 ms
`transpose(1000×1000)`	≈0 ms	≈0 ms	≈0 ms
custom fn (`mapParallel` / `map` / `np.vectorize`)	64 ms	5.9 ms	0.53 ms

MatD wins 8/9 scored operations vs NumPy — matmul is the exception, the pure-JVM fallback losing ~4× to NumPy's native OpenBLAS — and wins or ties all scored operations vs Breeze (benchBreeze geometric mean ~4.3× over its 7 scored ops). The v0.14.1 chunked multi-accumulator parallel reduction keeps sum/mean/std 4–6× ahead of NumPy 2.4.x's SIMD reductions. matmul vs Breeze is tied — both ran the same fallback backend here; with JNIBLAS loaded both call OpenBLAS at the same latency.

Linux (Intel Core i5-6500, Ubuntu 24.04, OpenBLAS)

Breeze/MatD: uni 0.14.1 / Scala 3.8.2 / JVM 21, both using native OpenBLAS via netlib JNIBLAS.

Operation	Breeze	MatD	Bz/MD
`randn(1000×1000)`	57.7 ms	15.6 ms	3.7×
`matmul 512×512`	1.83 ms	1.82 ms	1.00×
`sigmoid(1000×1000)`	9.52 ms	3.66 ms	2.6×
`relu(1000×1000)`	3.87 ms	1.06 ms	3.7×
`add(1000×1000)`	1.85 ms	2.34 ms	0.79×
`sum(1000×1000)`	1.17 ms	0.18 ms	6.7×
`mean(1000×1000)`	7.33 ms	0.27 ms	27×
`std(1000×1000)`	8.97 ms	1.55 ms	5.8×
`transpose(1000×1000)`	≈0 ms	≈0 ms	—
custom fn (`mapParallel` / `map`)	10.68 ms	1.08 ms	9.9×

MatD faster 8/9 scored, geometric mean 4.03× faster than Breeze. matmul is tied (both use OpenBLAS via JNIBLAS); add is the lone Breeze win (0.79×). sum flipped to a 6.7× MatD win after the v0.14.1 chunked parallel-reduction rewrite.

macOS (Apple Silicon)

Breeze/MatD: uni 0.14.1 / Scala 3.8.2 / JVM 21, both using native OpenBLAS via netlib JNIBLAS.

Operation	Breeze	MatD	Bz/MD
`randn(1000×1000)`	40.0 ms	3.93 ms	10.2×
`matmul 512×512`	1.06 ms	1.05 ms	1.01×
`sigmoid(1000×1000)`	12.5 ms	2.64 ms	4.7×
`relu(1000×1000)`	1.66 ms	0.37 ms	4.5×
`add(1000×1000)`	0.78 ms	1.04 ms	0.75×
`sum(1000×1000)`	0.97 ms	0.10 ms	9.3×
`mean(1000×1000)`	5.45 ms	0.08 ms	72×
`std(1000×1000)`	8.06 ms	1.03 ms	7.8×
`transpose(1000×1000)`	≈0 ms	≈0 ms	—
custom fn (`mapParallel` / `map`)	5.99 ms	0.42 ms	14.2×

MatD faster 8/9 scored, geometric mean 6.11× faster than Breeze. matmul is tied (both use OpenBLAS via JNIBLAS); add is the lone Breeze win (0.75×).

3PRF (Three-Pass Regression Filter)

Measured on Windows 11 (uni 0.15.1+ / JVM 22 / Scala 3.8.4, vs Python 3.13.13 / NumPy on scipy-openblas, and the t3prf Rust crate on OpenBLAS; medians of 25 timed calls per OOS procedure after explicit warm-up). See jsrc/tprf3Bench.sc, rust/src/bin/bench_tprf3.rs and Kelly & Pruitt (2015).

Through v0.14.1 both implementations were kept equivalently optimized, so published ratios compared equivalent algorithms: v0.14.0 rewrote the K&P J(k) centering products as O(T·N) centering on both sides, and v0.14.1 tuned both again — Python's three lstsq passes became normal-equations solves and its std/mean paths are NaN-gated; Scala's OOS windows gained fused, copy-free standardization.

v0.15.1 tuned the Scala side only, which briefly made the OOS rows a comparison between differently-optimized algorithms. Those three techniques are now applied to NumPy as well, so the OOS comparison is back to being like-for-like: both sides push the column scaling onto an (L+1)-column operand instead of materializing X·D⁻¹, run pass 2 in natural order, and derive each window's pass-1 cross products — and the per-window column std — by downdating full-sample ones. The same three are in the Rust port. For a Cross Val window dropping a single row, downdating turns pass 1 from O(T·N·L) into O(N·L).

NumPy keeps the direct per-window path where the downdate does not apply: autoproxy rebuilds the proxies each inner iteration, so there is nothing to precompute, and the downdates have no NaN-aware form. OOS Rolling is excluded for a different reason — its kept set is the short window itself, so a direct pass is already the cheap side.

Operation	Python	MatD	Rust	Py/MatD	MatD/Rust
`3PRF IS Full (T=650, N=40, L=2)`	0.27 ms	0.202 ms	0.051 ms	1.3× faster	4.0× faster
`3PRF OOS Recursive (T=650, N=40, L=2)`	21.7 ms	1.646 ms	1.42 ms	13.2× faster	1.2× faster
`3PRF OOS Cross Val (T=650, N=40, L=2)`	38.6 ms	5.735 ms	2.06 ms	6.7× faster	2.8× faster

Porting those optimizations to NumPy moved OOS Cross Val from 84.3 to 38.6 ms and OOS Recursive from 30.6 to 21.7 ms, which halved MatD's lead on Cross Val — it read 13.6× against the un-ported NumPy and reads 6.7× against the ported one. That difference is the measure of how much of the earlier gap was algorithm rather than language.

Python and MatD come from one run of uni.apps.Tprf3Bench, which measures both in the same process, so the Py/MatD ratios are internally consistent rather than spliced across sessions. The Rust column is from a --features blas build measured separately — see below for why that is the right baseline. That run's pure-Rust figures (3.54 / 4.51 ms on the two OOS rows) sit within 5% of the pure-Rust column this run produced (3.69 / 4.31 ms), which is the check that the two sessions are comparable at all.

Earlier releases published median-of-3 figures for the two-way comparison; those read 0.31 / 0.36 ms for IS Full, 31.8 / 1.8 ms for OOS Recursive and 88.9 / 7.1 ms for OOS Cross Val.

All three languages now run on bit-identical input matrices — each seeds a NumPy-compatible PCG64 with 0 and draws X, y, Z in that order, so the table measures implementations and not a mix of implementation and input. Until the Rust bench moved to NumPyRng it generated its own inputs from StdRng and a Box–Muller transform.

The Rust column is a --features blas build, which is the like-for-like comparison: MatD runs on JNIBLAS-backed native OpenBLAS (the benchmark log names it), so a default pure-Rust build is not the right baseline. Pure-Rust reads 0.052 / 3.69 / 4.31 ms on the same rows — roughly 2.3× slower on both OOS procedures, enough to turn OOS Recursive from a 1.2× win into a 2.2× loss, because those gemms are skinny (one operand has only L+1 columns) and matrixmultiply handles that shape poorly. It reverses at the small size, where BLAS call overhead outweighs the kernel win: at T=200/N=30 OOS Recursive is faster pure-Rust (0.38 vs 0.51 ms). Neither build dominates, so the benchmark prints its config: line and any quoted figure should say which build it is.

IS Full stays close — both sides reduce to the same two batch solves. Measured in isolation MatD runs 0.25 ms; the figure above comes from a full benchmark run, where it shares JIT call-site profiles with the OOS procedures.

The Linux and macOS tables below predate the Rust port and carry no Rust column; the crate has not been benchmarked on those platforms.

Linux (Intel Core i5-6500, Ubuntu 24.04, vs Python 3.12.3 / system OpenBLAS):

Operation	Python	MatD	Ratio
`3PRF IS Full (T=650, N=40, L=2)`	3.1 ms	0.46 ms	6.6× faster
`3PRF OOS Recursive (T=650, N=40, L=2)`	330 ms	47 ms	7.0× faster
`3PRF OOS Cross Val (T=650, N=40, L=2)`	862 ms	86 ms	10.0× faster

macOS (Apple Silicon, vs Python 3.14.6):

Operation	Python	MatD	Ratio
`3PRF IS Full (T=650, N=40, L=2)`	0.79 ms	0.22 ms	3.6× faster
`3PRF OOS Recursive (T=650, N=40, L=2)`	171 ms	22 ms	7.8× faster
`3PRF OOS Cross Val (T=650, N=40, L=2)`	447 ms	48 ms	9.2× faster

Label	Description
IS Full	In-sample fit over the full sample; two vectorized batch solves
OOS Recursive	Expanding-window out-of-sample; re-estimates at each step
OOS Cross Val	Leave-one-out cross-validation across all T windows

Full results and methodology: MatD Cheat Sheet — Performance.

Design Philosophy

uni.Mat is built on the principle that developers shouldn't have to choose between type safety and the ergonomics of NumPy. Its design leverages strides, offsets, and broadcasting for high efficiency—including LAPACK integration—all while maintaining a clean, expression-oriented API.

No element boxing — by construction, not by convention

A generic JVM matrix Mat[T] stores its data in an Array[T] that erases to Object[], so the natural m(i, j) element read and m(i, j) = v write would box every value into a java.lang.Double/java.lang.Float. uni avoids this without asking the client to learn a special "fast" accessor: for the two numeric workhorses, MatD (Mat[Double]) and MatF (Mat[Float]), the entire indexing family — scalar access and assignment, row/column/range slices, boolean masking, fancy Array[Int] indexing, and slice assignment — is supplied in a type-specialized form that reads and writes the primitive backing array directly.

So ordinary client code written the obvious way is already allocation-free for the element value:

import uni.data.*

val m = MatD((1.0, 2.0, 3.0), (4.0, 5.0, 6.0))
val x = m(0, 1)        // primitive double load — no java.lang.Double allocated
m(1, 2) = 99.0         // primitive double store
m(0, ::) = 0.0         // slice assignment, also unboxed

This is verified at the bytecode level (primitive daload/dastore, no Double.valueOf) and confirmed end-to-end by JFR allocation profiling, which shows zero boxed-scalar allocations on the element path. The MatF specialization is generated from the MatD source by the build, so the two never drift.

Note the scope: this covers the element value type on MatD/MatF. Other element types (Mat[Int], Mat[Big], …) still box element access — an inherent cost of generic JVM arrays — and a few collection-iteration paths box loop indices (Int), which is unrelated to the numeric data.

Test Coverage

Measured with sbt jacoco (1,980 tests). Branch coverage measures both true and false outcomes of every conditional, so it is typically lower than line coverage. Note: inline annotations were removed before running JaCoCo to prevent Scala 3's per-call-site bytecode duplication — where each call site gets its own branch counters — from artificially lowering reported coverage.

Package	Branch	Line
`uni.data`	70%	92%
`uni.stats`	32%	91%
`uni.io`	50%	83%
`uni.time`	34%	81%
`uni`	28%	60%
`uni.cli`	23%	59%

Installation

Add the following to your build.sbt:

libraryDependencies += "org.vastblue" %% "uni" % "0.14.1"

Native BLAS backend

uni uses netlib-java for matrix multiply. The native backend is selected automatically:

Platform	Backend	User action
macOS	Accelerate framework (always present)	None
Windows	VectorBLAS (Java Vector API SIMD)	None — see note below
Linux	`libblas.so.3` resolved by the system linker	See below

Linux: libblas.so.3 is a system symlink managed by update-alternatives. For maximum performance install an optimized BLAS and let the alternatives system point the symlink at it:

# Ubuntu / Debian — installs OpenBLAS and sets libblas.so.3 → libopenblas.so.3
sudo apt-get install libopenblas0

Without this, libblas.so.3 may resolve to the slow single-threaded reference BLAS. If libblas.so.3 is missing entirely, uni falls back to a pure-JVM BLAS (correct but slower).

Windows note (0.11.0): Native OpenBLAS support for Windows has merged in the netlib upstream but is pending publication (expected as 3.1.2). In the meantime, uni 0.11.0 uses VectorBLAS (Java Vector API SIMD, JDK 21+) for matrix multiply — correct and faster than scalar Java, but slower than native OpenBLAS for large matrices. A uni patch will update to 3.1.2 once it is published. Prior to 0.11.0, uni used bytedeco/openblas-platform which bundled native OpenBLAS for Windows automatically.

Advanced Usage

Quick Start: Mat Quick Start Guide — Fast track to NumPy-compatible matrix operations in Scala.
Visualization: Plot Guide — uni.plot methods, PlotStyle configuration, and demo scripts.
API Reference: Mat Reference Guide — Comprehensive API documentation with validated examples.
Cheat Sheet: MatD Cheat Sheet — Side-by-side comparison of MatD vs NumPy, Breeze, R, and MATLAB.
High Precision: Big Type Guide — High-precision matrices using MatB (Mat[Big]).
Path & I/O Reference: Path/String/File Extensions — Complete extension method reference for Path, JFile, and String.

NumPy to uni.MatD Mapping

NumPy	uni.MatD	Note
`a @ b`	`a *@ b`	Matrix multiplication
`a * b`	`a * b`	Element-wise product
`a[0, :]`	`a(0, ::)`	Row slice
`a[:, 0]`	`a(::, 0)`	Column slice
`np.apply_along_axis(f, 0, a)`	`a.eachCol.map(f)`	Apply f to each column
`np.apply_along_axis(f, 1, a)`	`a.eachRow.map(f)`	Apply f to each row
`a.reshape(r, c)`	`a.reshape(r, c)`	Change shape
`a.T`	`a.T`	$O(1)$ view
`np.random.randn(n)`	`MatD.randn(n)` `MatD.rnorm(n)`	n×1 column vector, standard normal
`np.random.randn(r,c)`	`MatD.randn(r, c)`	r×c matrix, standard normal
`np.where(c, x, y)`	`MatD.where(c, x, y)`	Conditional selection
`np.vstack(...)`	`MatD.vstack(...)`	Row-wise stack
`np.vsplit(m, n)`	`m.vsplit(n)`	Row-wise split
`m.item()`	`m.item`	Extract scalar from 1×1 matrix

Quick Start

Matrix Creation & Randomness

#!/usr/bin/env -S scala-cli shebang -Wunused:imports -Wunused:locals -deprecation

//> using dep org.vastblue:uni_3:0.14.2

import uni.data.*

// 100% faithful to NumPy's PCG64-based np.random.uniform
MatD.setSeed(42)
val weights = MatD.uniform(low = -0.1, high = 0.1, rows = 64, cols = 32)

// Standard constructors
val zeros    = MatD.zeros(10, 10)
val identity = MatD.eye(5)
val normal   = MatD.randn(10, 10)
println(normal)

Vector Types

CVec[T] (column vector, n×1) and RVec[T] (row vector, 1×n) are opaque types backed by Mat[T]. They add type-safe BLAS-style vector dispatch on top of Mat.

import uni.data.*

val y: CVecD = CVec(1.0, 2.0, 3.0)   // column vector
val r: RVecD = RVec(4.0, 5.0, 6.0)   // row vector

// Transpose flips column ↔ row
val rt: CVecD = r.T
val yt: RVecD = y.T

// *@ dispatch table
val dot:   Double = y.T *@ y          // RVec *@ CVec  → scalar
val dot2:  Double = y *@ y            // CVec *@ CVec  → scalar (auto-transpose)
val dot3:  Double = r *@ r            // RVec *@ RVec  → scalar (auto-transpose)
val outer: MatD   = y *@ y.T          // CVec *@ RVec  → n×n matrix
val Xy:    CVecD  = X *@ y            // Mat  *@ CVec  → CVec
val yTX:   RVecD  = y.T *@ X         // RVec *@ Mat   → RVec

// Arithmetic
val sum  = y + CVec(0.1, 0.2, 0.3)   // CVec + CVec
val yp1  = y + 1.0                    // CVec + scalar
val ym1  = y - 1.0                    // CVec - scalar
val s2   = 2.0 * y                    // Double * CVec
val si   = 2   * y                    // Int    * CVec
val sl   = 2L  * y                    // Long   * CVec

// Norm, show
val n: Double = y.norm
println(y.show)   // "3x1 CVec[Double]: ..."

CVec factory
`CVec(1.0, 2.0, 3.0)`	from varargs
`CVec.zeros[Double](n)`	n zeros
`CVec.ones[Double](n)`	n ones
`CVec.fromArray(arr)`	from `Array[T]`
`CVec.fromMat(m)`	from n×1 or 1×n `Mat[T]`

RVec has the same factory methods.

Matrix Type Aliases

To keep your code concise and idiomatic, uni.MatD provides type aliases and matching factory objects for supported numeric types.

Alias	Full Type	Description
`MatD`	`Mat[Double]`	Standard 64-bit floating point matrix (default)
`MatB`	`Mat[Big]`	High-precision, NaN-safe `BigDecimal` matrix
`MatF`	`Mat[Float]`	32-bit floating point matrix for memory efficiency
`CVecD`	`CVec[Double]`	Column vector, n×1
`RVecD`	`RVec[Double]`	Row vector, 1×n

Each alias has a matching factory object mirroring the MatD API:

#!/usr/bin/env -S scala-cli shebang -Wunused:imports -Wunused:locals -deprecation

//> using dep org.vastblue:uni_3:0.14.2

import uni.data.*

val weights:    MatD = MatD.randn(64, 32)
val prices:     MatB = MatB.zeros(10, 1)
val embeddings: MatF = MatF.rand(128, 64)

// Extension method from Big
val preciseVal = 10.5.asBig
val identityB: MatB = MatB.eye(5)

NumPy-like Syntax and Features

Slicing and Views

#!/usr/bin/env -S scala-cli shebang -Wunused:imports -Wunused:locals -deprecation

//> using dep org.vastblue:uni_3:0.14.2

import uni.data.*
import uni.data.MatD.*

MatD.setSeed(95)
val data = MatD.randn(100, 10)

// Extract a row or column as a view using the :: sentinel
val row = data(0, ::)    // first row
val col = data(::, 0)    // first column

// Constant-time transpose (no data copy)
val rotated = data.T
println(s"row: $row")
println(s"col: $col")
println(s"rotated: $rotated")
println(s"rotated: ${rotated.show("%7.2f")}")

Mathematical Operations

#!/usr/bin/env -S scala-cli shebang -Wunused:imports -Wunused:locals -deprecation

//> using dep org.vastblue:uni_3:0.14.2

import uni.data.*

val a = MatD.randn(3, 3)
val b = MatD.randn(3, 3)

val c = a *@ b    // matrix multiplication (matmul)
val d = a + b     // element-wise addition
val e = a * b     // Hadamard (element-wise) product
val f = a.relu    // built-in activation function

In-place Operations

#!/usr/bin/env -S scala-cli shebang -Wunused:imports -Wunused:locals -deprecation

//> using dep org.vastblue:uni_3:0.14.2

import uni.data.*
import uni.data.MatD.*

val m = MatD.ones(4, 4)
m :+= 2.0    // add scalar in-place
m :-= 1.0    // subtract scalar in-place
m :*= 3.0    // multiply scalar in-place
m :/= 2.0    // divide scalar in-place

val n = MatD.ones(4, 4)
m :+= n      // element-wise add matrix in-place

Boolean Operations

#!/usr/bin/env -S scala-cli shebang -Wunused:imports -Wunused:locals -deprecation

//> using dep org.vastblue:uni_3:0.14.2

import uni.data.*
import uni.data.MatD.*

val a    = MatD.randn(3, 3)
val mask = (a :== 0) || (a :== 1)    // MatD[Boolean]

val inverted = !mask
val count    = mask.sum    // count of true elements
val anyTrue  = mask.any
val allTrue  = mask.all

Stacking and Splitting

#!/usr/bin/env -S scala-cli shebang -Wunused:imports -Wunused:locals -deprecation

//> using dep org.vastblue:uni_3:0.14.2

import uni.data.*
import uni.data.MatD.*

val top    = MatD.ones(2, 4)
val bottom = MatD.zeros(2, 4)

val stacked = MatD.vstack(top, bottom)    // 4x4 MatD
val halves  = stacked.vsplit(2)          // Seq of two 2x4 Mats

val left  = MatD.ones(4, 2)
val right = MatD.zeros(4, 2)
val wide  = MatD.hstack(left, right)      // 4x4 MatD
val cols  = wide.hsplit(2)               // Seq of two 4x2 Mats

Display and Formatting

#!/usr/bin/env -S scala-cli shebang -Wunused:imports -Wunused:locals -deprecation

//> using dep org.vastblue:uni_3:0.14.2

import uni.data.*
import uni.data.MatD.*

val m = MatD.randn(3, 3)

println(m)               // calls toString
println(m.show)          // equivalent, explicit
println(m.show("%.2f"))  // custom format string

// Adjust truncation thresholds for large matrices
MatD.setPrintOptions(maxRows = 20, maxCols = 20, edgeItems = 5)

Usage Example

Comprehensive Example: MatD in Action

The following example demonstrates a wide array of uni.MatD capabilities , including matrix creation, slicing, broadcasting, linear algebra (SVD, QR, Inverse), and in-place mutation. It illustrates how the library brings NumPy-style ergonomics to Scala while maintaining high performance.

#!/usr/bin/env -S scala-cli shebang -Wunused:imports -Wunused:locals -deprecation
//> using dep org.vastblue:uni_3:0.14.2

import uni.data.*
import uni.data.MatD.*

object MatDCheck {
  def main(args: Array[String]): Unit = {
    // 1. Creation and Basic Shapes
    val m = MatD.zeros[Double](3, 4)
    val v = MatD.row(1, 2, 3, 4)
    val eye = MatD.eye(2)
    val r = MatD.arange(0, 10, 2)
    val ls = MatD.linspace(0, 1, 5)

    // 2. Arithmetic & Broadcasting
    val result = m + v      // Row-wise broadcasting
    val powr = result ~^ 2  // Power operator
    val matMult = powr *@ v.T // Matrix multiplication (matmul)

    // 3. Math Functions
    val sq = m.sqrt
    val ex = v.exp
    val cl = v.clip(0, 5)

    // 4. Reductions
    val s = m.sum
    val s0 = m.sum(0) // Sum over axis 0
    val mn = m.mean

    // 5. Linear Algebra
    val A = MatD((1, 2), (3, 4))
    val b = MatD.row(5, 6)
    val x = A.solve(b)         // Linear solver
    val inv_A = A.inverse
    val det_A = A.determinant
    val (u, s2, vt) = A.svd    // Singular Value Decomposition
    val (q, r2) = A.qrDecomposition

    // 6. Boolean Masking & Filtering
    val mask = v.gt(1)
    val filtered = v(mask)
    v(v.lt(0)) = 0 // Conditional assignment

    // 7. Slicing & Manipulation
    val sub = m(0 until 2, ::) // Slice rows
    val col = m(::, 0)         // Slice column
    val reshaped = m.reshape(2, 6)
    val transposed = m.T       // O(1) Transpose
    val stacked = MatD.vstack(m, m)

    // 8. Random Generation
    val rand_m = MatD.rand(3, 3)
    val randn_m = MatD.randn(3, 3)

    // 9. In-place Mutation
    m :+= 1
    m :*= 2
    for i <- 0 until 3 do
      m(i, ::) = i * 2
  }
}

Example: Neural Network Layer

Because activation functions are members of the MatD type, building layers is idiomatic:

#!/usr/bin/env -S scala-cli shebang -Wunused:imports -Wunused:locals -deprecation

//> using dep org.vastblue:uni_3:0.14.2

import uni.data.*

def denseLayer(input: MatD, weights: MatD, bias: MatD): MatD = {
  // Simple, readable forward pass
  (input *@ weights + bias).sigmoid
}

Other Features

uni also includes utilities that predate uni.data.MatD and remain available via import uni.*:

Scripting Shortcuts: Portable Programming Utilities — MSYS2/Cygwin-aware paths, smart date parsing, command-line argument handling, and inline data embedding.
Smart Date-Time Parsing: Date-Time Parser Guide — Autodetect messy dates, configure MDY/DMY preferences, and eliminate ETL regex steps.

Tolerant Numeric Parsing for Big Data Prep

Raw financial and scientific datasets rarely arrive in clean form. uni.data.BigUtils provides isNumeric and getMostSpecificType that accept the messy formats found in spreadsheets and CSV exports without requiring pre-cleaning:

Raw input	Recognised as
`"$1,234.56"`	`Big(1234.56)`
`"(500.00)"`	`Big(-500.00)` — accounting negative
`"7.5K"` / `"2.1M"` / `"4B"`	`Big(7500)` / `Big(2_100_000)` / `Big(4_000_000_000)`
`"12.5%"`	`Big(0.125)` — already divided by 100
`"-0.042"`	`Big(-0.042)`
`"2024-01-15"`	`DateTime`
anything else	`String` (passed through unchanged)

getMostSpecificType promotes each raw cell to the most specific type it can without error, so a column of mixed-format numbers loads correctly as Mat[Big] in a single pass:

#!/usr/bin/env -S scala-cli shebang -Wunused:imports -Wunused:locals -deprecation

//> using dep org.vastblue:uni_3:0.14.2

import uni.data.*
import uni.data.BigUtils.*

val raw = Seq("$1,200", "(300)", "1.5K", "0.75%", "N/A", "2024-03-01")
val typed = raw.map(getMostSpecificType)
// => Seq(Big(1200), Big(-300), Big(1500), Big(0.0075), "N/A", DateTime(...))

// Load numeric cells directly into a column vector
val nums = typed.collect { case b: BigDecimal => Big(b) }
val col  = MatB.col(nums*)

This eliminates the typical ETL step of writing format-specific regex cleaners before ingestion — particularly valuable when working with multi-source datasets where currency symbols, parenthesised negatives, and scale suffixes appear unpredictably across columns.

Name		Name	Last commit message	Last commit date
Latest commit History 144 Commits
.github		.github
data		data
datasets		datasets
doc		doc
docs		docs
images		images
jsrc		jsrc
project		project
py		py
rust		rust
src		src
t3prf-validation		t3prf-validation
test-data/tprf3-parity		test-data/tprf3-parity
.gitattributes		.gitattributes
.gitignore		.gitignore
.jvmopts		.jvmopts
.sbtopts		.sbtopts
.scalafmt.conf		.scalafmt.conf
CHANGELOG.md		CHANGELOG.md
LICENSE		LICENSE
README.md		README.md
build.sbt		build.sbt
concept.txt		concept.txt
release-and-publish.sh		release-and-publish.sh

Navigation Menu

Repository files navigation

uni.data

Key Features

Rust companion crate (`rust/`)

Visualization (`uni.plot`)

Performance

Core matrix operations

Linux (Intel Core i5-6500, Ubuntu 24.04, OpenBLAS)

macOS (Apple Silicon)

3PRF (Three-Pass Regression Filter)

Design Philosophy

No element boxing — by construction, not by convention

Test Coverage

Installation

Native BLAS backend

Advanced Usage

NumPy to uni.MatD Mapping

Quick Start

Matrix Creation & Randomness

Vector Types

Matrix Type Aliases

NumPy-like Syntax and Features

Slicing and Views

Mathematical Operations

In-place Operations

Boolean Operations

Stacking and Splitting

Display and Formatting

Usage Example

Comprehensive Example: MatD in Action

Example: Neural Network Layer

Other Features

Tolerant Numeric Parsing for Big Data Prep

About

Releases

Packages

Used by

Contributors

Languages

Folders and files

Latest commit

History

Repository files navigation

uni.data

Key Features

Rust companion crate (rust/)

Visualization (uni.plot)

Performance

Core matrix operations

Linux (Intel Core i5-6500, Ubuntu 24.04, OpenBLAS)

macOS (Apple Silicon)

3PRF (Three-Pass Regression Filter)

Design Philosophy

No element boxing — by construction, not by convention

Test Coverage

Installation

Native BLAS backend

Advanced Usage

NumPy to uni.MatD Mapping

Quick Start

Matrix Creation & Randomness

Vector Types

Matrix Type Aliases

NumPy-like Syntax and Features

Slicing and Views

Mathematical Operations

In-place Operations

Boolean Operations

Stacking and Splitting

Display and Formatting

Usage Example

Comprehensive Example: MatD in Action

Example: Neural Network Layer

Other Features

Tolerant Numeric Parsing for Big Data Prep

About

Topics

Resources

Stars

Watchers

Forks

Releases

Packages

Used by

Contributors

Languages

Rust companion crate (`rust/`)

Visualization (`uni.plot`)