Julienne

Julienne: Idiomatic Correctness Checking for Fortran 2023

The Julienne framework offers a unified approach to unit testing and runtime assertion checking. Julienne defines idioms for specifying correctness conditions in a common way when writing tests that wrap the tested procedures or assertions that conditionally execute inside procedures to check correctness. Julienne's idioms center around expressions built from defined operations: a uniquely flexible Fortran capability allowing developers to define new operators in addition to overloading Fortran's intrinsic operators. The following table provides examples of the expressions Julienne supports:

where * .isAtLeast. and .isAtMost. can alternatively be spelled .greaterThanOrEqualTo. and .lessThanOrEqualTo., respectively; * .equalsExpected. uses ==, which implies that trailing blank spaces are ignored in character operands;
* .equalsExpected. with integer operands supports default integers and integer(c_size_t); * .isBefore. and .isAfter. verify alphabetical and reverse-alphabetical order, respectively; and * .all. aggregates arrays of expression results, reports a consensus result, and shows diagnostics only for failing tests, if any; * .equalsExpected. generates asymmetric diagnostic output for failures, denoting the left- and right-hand sides as the actual value and expected values, respectively; and. * // appends the subsequent string to diagnostics strings, if any.

Expressive idioms

Assertions

Any of the above expressions can be the actual argument in an invocation of Julienne's call_julienne_assert function-line preprocessor macro:

call_julienne_assert(x .lessThan. y)

which a preprocessor will replace with a call to Julienne's assertion subroutine when compiling with -DASSERTIONS. Otherwise, the preprocessor will remove the above line entirely when -DASSERTIONS is not present.

Unit tests

The above tabulated expressions can also serve as function results in unit tests.

Constraints

All operands in an expression must be compatible in type and kind as well as conformable in rank, where the latter condition implies that the operands must be all scalars or all arrays with the same shape or a combination of scalars and arrays with the same shape. This constraint follows from each of the binary operators being elemental. The unary .all. operator applies to operands of any rank.

Each tabulated expression above produces a test_diagnosis_t object with two components:

• a logical indicator of test success if .true. or failure if .false. and
• an automated diagnostic messages generated only if the test or assertion fails.

Custom Test Diagnostics

For cases in which Julienne's operators do not support the desired correctness condition, the framework provides string-handling utilities for use in crafting custom diagnostic messages. The string utilities center around Julienne's string_t derived type, which offers elemental constructor functions, i.e., functions that one invokes via the same name as the derived type: string_t(). The string_t() constructor functions convert data of numeric type to character type, storing the resulting character representation in a private component of the constructor function result. The actual argument provided to the constructor function can be of any one of several types, kinds, and ranks.

Julienne provides defined operations for concatenating string_t objects (//), forming a concatenated string_t object from an array of string_t objects (.cat.), forming a separated-value list (.separatedBy. or equivalently .sv.), including a comma-separated value list (.csv.). The table below shows some expressions that Julienne supports with these defined operations.

One can use such expressions to craft a diagnostic message when constructing a custom test function result:

type(test_diagnosis_t) test_diagnosis
test_diagnosis = test_diagnosis_t( &
  test_passed = i==j, &
  diagnostics_string = "expected " // string_t(i) // "; actual " //string_t(j) &
)

A file abstraction

Arrays of string_t objects provide a convenient way to store a ragged-length array of character data. Julienne's file_t derived type has a private component that is a string_t array, wherein each element is one line of a text file. By storing a file in a file_t object using the file_t derived type's constructor function one can confine a program's file input/output (I/O) to one or two procedures. The resulting file_t object can be manipulated elsewhere without incurring the costs associated with file I/O. For example, the following line reads a file named data.txt into a file_t object and associates the name file with the resulting object.

type(file_t) file
associate(file => file_t("data.txt"))
end associate

This style supports functional programming patterns in two ways. First, the rest of the program can be comprised of pure procedures, which are precluded from performing I/O. Second, an associate name is immutable when associated with an expression, including an expression that is simply a function reference. Functional programming revolves around creating and using immutable state. (By contrast, when associating a name with a variable or array instead of with an expression, only certain attributes, such as the entity's allocation status, are immutable. The value of such a variable or array can be redefined.)

Functional Programming

Functional programming patterns centered around pure procedures enhance code clarity, ease refactoring, and encourage optimization. For example, the constraints on pure procedures make it easier for a developer or a compiler to safely reorder program statements. Moreover, Fortran allows invoking only pure procedures inside do concurrent, a construct that compilers can automatically offload to a graphics processing unit (GPU).

Julienne lowers a widely stated barrier to writing pure procedures (including simple procedures): the difficulty in printing values while debugging code. The Julienne philosophy is that printing a value for debugging purposes implies an expectation about the value. Assert such expectations by writing Julienne expressions inspired by natural language. A program will proceed quietly past a correct assertion. An incorrect assertion produces either automated or custom diagnostic messages during error termination.

Getting Started

Writing Unit Tests

Please see demo/README.md for a detailed demonstration of test setup.

Writing Assertions

To write a Julienne assertion, insert a function-like preprocessor macro call_julienne_assert on a single line as in each of the two macro invocations below:

#include "julienne-assertion-macros.h"
program main
  use, julienne_m, only : call_julienne_assert_
  implicit none
  real, parameter :: x=1., y=2., tolerance=3.
  call_julienne_assert(x .approximates. y .within. tolerance)
  call_julienne_assert(abs(x-y) < tolerance)
end program

where inserting -DASSERTIONS in a compile command will expand the macros to

  call call_julienne_assert_(x .approximates. y .within. tolerance, __FILE__, __LINE__)
  call call_julienne_assert_(allocated(a), __FILE__, __LINE__)

and where dots (.) delimit Julienne operators. The above expression containing Julienne operators evaluates to a Julienne test_diagnosis_t object, whereas expression allocated(a) on the subsequent line evaluates to a logical value. If an assertion containing a Julienne expression fails, Julienne inserts diagnostic information into the stop code in an ultimate error stop. If an expression evaluates to a logical value of false., the error stop code will contain a literal copy of the expression (e.g., allocated(a)). In either case, Julienne also inserts the file and line number into the stop code using via the __FILE__ and __LINE__ macros, respectively. Most compilers write the resulting stop code to error_unit.

An Origin Story

Julienne's name derives from the term for vegetables sliced into thin strings: julienned vegetables. The Veggies and Garden unit-testing frameworks inspired the structure of Julienne's tests and output. Initially developed in the Sourcery repository as lightweight alternative with greater portability across compilers, Julienne's chief innovation now lies in the expressive idioms the framework supports.

Building and Testing

Compiler support

When built with the compiler versions tabulated below, all Julienne tests pass.

With gfortran 13 through 14.2.0, - The test_description_t constructor's diagnosis_function actual argument must be a procedure pointer declared with procedure(diagnosis_function_i). - The string_t type-bound function bracket crashes.

Build/test commands

LLVM (flang-new) compiler

With version 20 or later, please run

fpm test --compiler flang-new --flag "-O3"

With version 19, please run

fpm test --compiler flang-new --flag "-O3 -mmlir -allow-assumed-rank"

to enable support for assumed-rank dummy arguments.

NAG (nagfor) compiler

Serial execution

fpm test --compiler nagfor --flag "-fpp -O3"

Parallel execution

To execute build and run parallel tests in 2 images, please run

export NAGFORTRAN_NUM_IMAGES=2
fpm test --compiler nagfor --flag "-fpp -O3 -coarray"

Replace the "2" above with any number up to the compiler limit of 1000 images.

GNU (gfortran) compiler

Serial execution

For compiler versions 14 or higher, plese use

fpm test --compiler gfortran --profile release

For version 13, please append --flag "-ffree-line-length-none" to the above command to enable the Fortran 2023 line-length maximum of 5000 characters.

Parallel execution

With OpenCoarrays installed, please replace --compiler gfortran with --compiler caf and please append --runner "cafrun -n 2" to the above command.

Intel (ifx) compiler

Serial execution

fpm test --compiler ifx --flag "-fpp -O3" --profile release

Parallel execution

To execute build and run parallel tests in 2 images, please run

export FOR_COARRAY_NUM_IMAGES=2
fpm test --compiler ifx --flag "-fpp -O3 -coarray" --profile release

Replace the "2" above with any desired number of images.

Documentation

See our online documentation or build the documentation locally by installing FORD and executing ford ford.md.

Known Software Using Julienne

• Fiats: Functional inference and training for surrogates
• Matcha: Motility analysis of T-cell histories in activation
• TRACE two-phase flow solver for nuclear reactors
• nQMCC: Quantum Monte Carlo simulation software for nuclear physics