coffi

Coffi is a foreign function interface library for Clojure, using the new Project Panama that's a part of the preview in Java 19. This allows calling native code directly from Clojure without the need for either Java or native code specific to the library, as e.g. the JNI does. Coffi focuses on ease of use, including functions and macros for creating wrappers to allow the resulting native functions to act just like Clojure ones, however this doesn't remove the ability to write systems which minimize the cost of marshaling data and optimize for performance, to make use of the low-level access Panama gives us.

Installation

This library is available on Clojars. Add one of the following entries to the :deps key of your deps.edn:

org.suskalo/coffi {:mvn/version "0.6.409"}
io.github.IGJoshua/coffi {:git/tag "v0.6.409" :git/sha "f974446"}

If you use this library as a git dependency, you will need to prepare the library.

$ clj -X:deps prep

Coffi requires usage of the package java.lang.foreign, and everything in this package is considered to be a preview release, which are disabled by default. In order to use coffi, add the following JVM arguments to your application.

--enable-preview --enable-native-access=ALL-UNNAMED

You can specify JVM arguments in a particular invocation of the Clojure CLI with the -J flag like so:

clj -J--enable-preview -J--enable-native-access=ALL-UNNAMED

You can also specify them in an alias in your deps.edn file under the :jvm-opts key (see the next example) and then invoking the CLI with that alias using -M, -A, or -X.

{:aliases {:dev {:jvm-opts ["--enable-preview" "--enable-native-access=ALL-UNNAMED"]}}}

Other build tools should provide similar functionality if you check their documentation.

Coffi also includes support for the linter clj-kondo. If you use clj-kondo and this library's macros are not linting correctly, you may need to install the config bundled with the library. You can do so with the following shell command:

$ clj-kondo --copy-configs --dependencies --lint "$(clojure -Spath)"

Usage

There are two major components to coffi and interacting with native code: manipulating off-heap memory, and loading native code for use with Clojure.

In the simplest cases, the native functions you call will work exclusively with built-in types, for example the function strlen from libc.

(require '[coffi.mem :as mem :refer [defalias]])
(require '[coffi.ffi :as ffi :refer [defcfn]])

(defcfn strlen
  "Given a string, measures its length in bytes."
  strlen [::mem/c-string] ::mem/long)

(strlen "hello")
;; => 5

The first argument to defcfn is the name of the Clojure var that will hold the native function reference, followed by an optional docstring and attribute map, then the C function identifier, including the name of the native symbol, a vector of argument types, and the return type.

If you wish to use a native function as an anonymous function, it can be done with the cfn function.

((ffi/cfn "strlen" [::mem/c-string] ::mem/long) "hello")
;; => 5

If you want to use functions from libraries other than libc, then you'll need to load them. Two functions are provided for this, load-system-library, and load-library. load-system-library takes a string which represents the name of a library that should be loaded via system lookup.

(ffi/load-system-library "z")

This will load libz from the appropriate place on the user's load path.

Alternatively, load-library takes a file path to a dynamically loaded library.

(ffi/load-library "lib/libz.so")

This will load libz from the lib subdirectory of the current working directory. As you can see this requires the entire filename, including platform-specific file extensions.

If a library is attempted to be loaded but doesn't exist or otherwise can't be loaded, an exception is thrown. This can be convenient as any namespace with a load-library call at the top level cannot be required without the library being able to be loaded.

Primitive Types

Coffi defines a basic set of primitive types:

byte
short
int
long
char
float
double
pointer

Each of these types maps to their C counterpart. Values of any of these primitive types except for pointer will be cast with their corresponding Clojure function when they are passed as arguments to native functions. Additionally, the c-string type is defined, although it is not primitive.

Composite Types

In addition, some composite types are also defined in coffi, including struct and union types (unions will be discussed with serialization and deserialization). For an example C struct and function:

typedef struct point {
    float x;
    float y;
} Point;

Point zero(void) {
    Point res = {};

    res.x = 0.0;
    res.y = 0.0;

    return res;
}

The corresponding coffi definition is like so:

(defcfn zero-point
  "zero" [] [::mem/struct [[:x ::mem/float] [:y ::mem/float]]])

(zero-point)
;; => {:x 0.0,
;;     :y 0.0}

Writing out struct definitions like this every time would get tedious, so the macro defalias is used to define a struct alias.

(defalias ::point
  [::mem/struct
   [[:x ::mem/float]
    [:y ::mem/float]]])

(defcfn zero-point
  "zero" [] ::point)

Struct definitions do not include any padding by default. Functions for transforming struct types to include padding conforming to various standards can be found in coffi.layout.

(require '[coffi.layout :as layout])

(defalias ::needs-padding
  (layout/with-c-layout
   [::mem/struct
    [[:a ::mem/char]
     [:x ::mem/float]]]))

(mem/size-of ::needs-padding))
;; => 8

(mem/align-of ::needs-padding)
;; => 4

Values deserialized with types produced from layout functions may include an extra :coffi.layout/padding key with a nil value.

A limitation of the defcfn macro in its current form is that types provided to it must be provided in a literal form, not as an expression that evaluates to a type. This means that if you wish to use a layout function on a struct you must define an alias for it before the type can be used as a type in defcfn.

In cases where a pointer to some data is required to pass as an argument to a native function, but doesn't need to be read back in, the pointer primitive type can take a type argument.

[::mem/pointer ::mem/int]

Arrays are also supported via a type argument. Keep in mind that they are the array itself, and not a pointer to the array like you might see in certain cases in C.

[::mem/array ::mem/int 3]

Callbacks

In addition to these composite types, there is also support for Clojure functions.

[::ffi/fn [::mem/c-string] ::mem/int]

Be aware though that if an exception is thrown out of a callback that is called from C, the JVM will crash. The resulting crash log should include the exception type and message in the registers section, but it's important to be aware of all the same. Ideally you should test your callbacks before actually passing them to native code.

When writing a wrapper library for a C library, it may be a good choice to wrap all passed Clojure functions in an additional function which catches all throwables, potentially notifies the user in some manner (e.g. logging), and returns a default value. This is on the wrapper library's developer to decide when and where this is appropriate, as in some cases no reasonable default return value can be determined and it is most sensible to simply crash the JVM. This is the reason that coffi defaults to this behavior, as in the author's opinion it is better to fail hard and fast rather than to attempt to produce a default and cause unexpected behavior later.

Another important thing to keep in mind is the expected lifetime of the function that you pass to native code. For example it is perfectly fine to pass an anonymous function to a native function if the callback will never be called again once the native function returns. If however it saves the callback for later use the JVM may collect it prematurely, causing a crash when the callback is later called by native code.

Variadic Functions

Some native functions can take any number of arguments, and in these cases coffi provides vacfn-factory (for "varargs C function factory").

(def printf-factory (ffi/vacfn-factory "printf" [::mem/c-string] ::mem/int))

This returns a function of the types of the rest of the arguments which itself returns a native function wrapper.

(def print-int (printf-factory ::mem/int))

(print-int "Some integer: %d\n" 5)
;; Some integer: 5

At the moment there is no equivalent to defcfn for varargs functions.

Some native functions that are variadic use the type va_list to make it easier for other languages to call them in their FFI. At the time of writing, coffi does not support va-list, however it is a planned feature.

Global Variables

Some libraries include global variables or constants accessible through symbols. To start with, constant values stored in symbols can be fetched with const, or the parallel macro defconst

(def some-const (ffi/const "some_const" ::mem/int))
(ffi/defconst some-const "some_const" ::mem/int)

This value is fetched once when you call const and is turned into a Clojure value. If you need to refer to a global variable, then static-variable (or parallel defvar) can be used to create a reference to the native value.

(def some-var (ffi/static-variable "some_var" ::mem/int))
(ffi/defvar some-var "some_var" ::mem/int)

This variable is an IDeref. Each time you dereference it, the value will be deserialized from the native memory and returned. Additional functions are provided for mutating the variable.

(ffi/freset! some-var 5)
;; => 5
@some-var
;; => 5

Be aware however that there is no synchronization on these types. The value being read is not read atomically, so you may see an inconsistent state if the value is being mutated on another thread.

A parallel function fswap! is also provided, but it does not provide any atomic semantics either.

The memory that backs the static variable can be fetched with the function static-variable-segment, which can be used to pass a pointer to the static variable to native functions that require it.

Complex Wrappers

Some functions require more complex code to map nicely to a Clojure function. The defcfn macro provides facilities to wrap the native function with some Clojure code to make this easier.

(defcfn takes-array
  "takes_array_with_count" [::mem/pointer ::mem/long] ::mem/void
  native-fn
  [ints]
  (let [arr-len (count ints)
        int-array (serialize ints [::mem/array ::mem/int arr-len])]
    (native-fn (mem/address-of int-array) arr-len)))

The symbol native-fn can be any unqualified symbol, and names the native function being wrapped. It must be called in the function body below if you want to call the native code.

This serialize function has a paired deserialize, and allows marshaling Clojure data back and forth to native data structures.

This can be used to implement out variables often seen in native code.

(defcfn out-int
  "out_int" [::mem/pointer] ::mem/void
  native-fn
  [i]
  (let [int-ptr (serialize i [::mem/pointer ::mem/int])]
    (native-fn int-ptr)
    (deserialize int-ptr [::mem/pointer ::mem/int])))

Sessions

Before JDK 19 Sessions were called Scopes. Coffi retains functions that are named for creating scopes for backwards compatibility, but they will be removed in version 1.0.

In order to serialize any non-primitive type (such as the previous [::mem/pointer ::mem/int]), off-heap memory needs to be allocated. When memory is allocated inside the JVM, the memory is associated with a session. Because none was provided here, the session is an implicit session, and the memory will be freed when the serialized object is garbage collected.

In many cases this is not desirable, because the memory is not freed in a deterministic manner, causing garbage collection pauses to become longer, as well as changing allocation performance. Instead of an implicit session, there are other kinds of sessions as well. A stack-session is a thread-local session. Stack sessions are Closeable, which means they should usually be used in a with-open form. When a stack-session is closed, it immediately frees all the memory associated with it. The previous example, out-int, can be implemented with a stack session.

(defcfn out-int
  "out_int" [::mem/pointer] ::mem/void
  native-fn
  [i]
  (with-open [session (mem/stack-session)]
    (let [int-ptr (mem/serialize i [::mem/pointer ::mem/int] session)]
      (native-fn int-ptr)
      (mem/deserialize int-ptr [::mem/pointer ::mem/int]))))

This will free the pointer immediately upon leaving the function.

When memory needs to be accessible from multiple threads, there's shared-session. When using a shared-session, it should be accessed inside a with-acquired block. When a shared-session is .closed, it will release all its associated memory when every with-acquired block associated with it is exited.

In addition, two non-Closeable sessions are global-session, which never frees the resources associated with it, and connected-session, which is a session that frees its resources on garbage collection, like an implicit session.

Serialization and Deserialization

Custom serializers and deserializers may also be created. This is done using two sets of three multimethods which can be extended by the user. For any given type, only one set need be implemented.

Two examples of custom types are given here, one is a 3d vector, and the other an example of a tagged union.

Vector3

For the vector type, it will serialize to a pointer to an array of three floats.

The multimethod primitive-type returns the primitive type that a given type serializes to. For this example, it should be a pointer.

(defmethod mem/primitive-type ::vector
  [_type]
  ::mem/pointer)

For any type which doesn't serialize to a primitive, it returns nil, and therefore need not be overriden.

Next is serialize* and deserialize*, multimethods that work with types that serialize to primitives.

(defmethod mem/serialize* ::vector
  [obj _type session]
  (mem/address-of (mem/serialize obj [::mem/array ::mem/float 3] session)))

(defmethod mem/deserialize* ::vector
  [addr _type]
  (mem/deserialize (mem/slice-global addr (mem/size-of [::mem/array ::mem/float 3]))
                   [::mem/array ::mem/float 3]))

The slice-global function allows you to take an address without an associated session and get a memory segment which can be deserialized.

In cases like this where we don't know the session of the pointer, we could use add-close-action! to ensure it's freed. For example if a free-vector! function that takes a pointer exists, we could use this:

(defcfn returns-vector
  "returns_vector" [] ::mem/pointer
  native-fn
  [session]
  (let [ret-ptr (native-fn)]
    (add-close-action! session #(free-vector! ret-ptr))
    (deserialize ret-ptr ::vector)))

This function takes a session and returns the deserialized vector, and it will free the pointer when the session closes.

Tagged Union

For the tagged union type, we will represent the value as a vector of a keyword naming the tag and the value. The type itself will need to take arguments, similar to struct. For example, if we were to represent a result type like in Rust, we might have the following values:

[:ok 5]
[:err "Invalid number format"]

To represent this, we can have a tagged-union type. For this instance of the result type, it may look like this:

[::tagged-union [:ok :err] {:ok ::mem/int :err ::mem/c-string}]

The native representation of these objects is a struct of the tag and a union of the value. In order to correctly serialize the data and pass it to native code, we need a representation of the native layout of the data. The c-layout multimethod provides that.

(defmethod mem/c-layout ::tagged-union
  [[_tagged-union tags type-map]]
  (mem/c-layout [::mem/struct
                 [[:tag ::mem/long]
                  [:value [::mem/union (vals type-map)]]]]))

Types with type arguments are represented as vectors of the type name and any additional arguments. The type name is what is dispatched on for the multimethods.

Now that we have a native layout, we need to be able to serialize and deserialize the value into and out of memory segments. This is accomplished with serialize-into and deserialize-from.

(defn item-index
  "Gets the index of the first occurance of `item` in `coll`."
  [coll item]
  (first
   (->> coll
        (map-indexed vector)
        (filter (comp #{item} second))
        (map first))))

(defmethod mem/serialize-into ::tagged-union
  [obj [_tagged-union tags type-map] segment session]
  (mem/serialize-into
   {:tag (item-index tags (first obj))
    :value (second obj)}
   [::mem/struct
    [[:tag ::mem/long]
     [:value (get type-map (first obj))]]]
   segment
   session))

This serialization method is rather simple, it just turns the vector value into a map, and serializes it as a struct, choosing the type of the value based on the tag.

(defmethod mem/deserialize-from ::tagged-union
  [segment [_tagged-union tags type-map]]
  (let [tag (mem/deserialize-from segment ::mem/long)]
    [(nth tags tag)
     (mem/deserialize-from
      (mem/slice segment (mem/size-of ::mem/long))
      (get type-map tag))]))

Deserialization is a little more complex. First the tag is retrieved from the beginning of the segment, and then the type of the value is decided based on that before it is deserialized.

Unions

In the last section the custom serialization and deserialization of a tagged union used a union from coffi in order to define the native layout, but not for actual serialization or deserialization. This is intentional. A union in coffi is rather limited. It can be serialized, but not deserialized without external information.

[::mem/union
 #{::mem/float ::mem/double}
 :dispatch #(cond
             (float? %) ::mem/float
             (double? %) ::mem/double)]

This is a minimal union in coffi. If the :dispatch keyword argument is not passed, then the union cannot be serialized, as coffi would not know which type to serialize the values as. In the example with a tagged union, a dispatch function was not provided because the type was only used for the native layout.

In addition to a dispatch function, when serializing a union an extract function may also be provided. In the case of the value in the tagged union from before, it could be represented for serialization purposes like so:

[::mem/union
 #{::mem/int ::mem/c-string}
 :dispatch #(case (first %)
              :ok ::mem/int
              :err ::mem/c-string)
 :extract second]

This union however would not include the tag when serialized.

If a union is deserialized, then all that coffi does is to allocate a new segment of the appropriate size with an implicit session so that it may later be garbage collected, and copies the data from the source segment into it. It's up to the user to call deserialize-from on that segment with the appropriate type.

Unwrapped Native Handles

Some native libraries work with handles to large amounts of data at once, making it undesirable to marshal data back and forth from Clojure, both because it's not necessary to work with the data in Clojure directly, or also because of the high (de)serialization costs associated with marshaling. In cases like these, unwrapped native handles are desirable.

The functions make-downcall and make-varargs-factory are also provided to create raw function handles.

(def raw-strlen (ffi/make-downcall "strlen" [::mem/c-string] ::mem/long))
(raw-strlen (mem/serialize "hello" ::mem/c-string))
;; => 5

With raw handles, the argument types are expected to exactly match the types expected by the native function. For primitive types, those are primitives. For addresses, that is MemoryAddress, and for composite types like structs and unions, that is MemorySegment. Both MemoryAddress and MemorySegment come from the java.lang.foreign package.

In addition, when a raw handle returns a composite type represented with a MemorySegment, it requires an additional first argument, a SegmentAllocator, which can be acquired with session-allocator to get one associated with a specific session. The returned value will live until that session is released.

In addition, function types can be specified as being raw, in the following manner:

[::ffi/fn [::mem/int] ::mem/int :raw-fn? true]

Clojure functions serialized to this type will have their arguments and return value exactly match the types specified and will not perform any serialization or deserialization at their boundaries.

One important caveat to consider when writing wrappers for performance-sensitive functions is that the convenience macro defcfn that coffi provides will already perform no serialization or deserialization on primitive arguments and return types, so for functions with only primitive argument and return types there is no performance reason to choose unwrapped native handles over the convenience macro.

Manual (De)Serialization

Coffi uses multimethods to dispatch to (de)serialization functions to enable code that's generic over the types it operates on. However, in cases where you know the exact types that you will be (de)serializing and the multimethod dispatch overhead is too high a cost, it may be appropriate to manually handle (de)serializing data. This will often be done paired with Unwrapped Native Handles.

Convenience functions are provided to both read and write all primitive types and addresses, including byte order.

As an example, when wrapping a function that returns an array of big-endian floats, the following code might be used.

(def ^:private returns-float-array* (ffi/make-downcall "returns_float_array" [::mem/pointer] ::mem/int))
(def ^:private release-floats* (ffi/make-downcall "releases_float_array" [::mem/pointer] ::mem/void))

(defn returns-float-array
  []
  (with-open [session (mem/stack-session)]
    (let [out-floats (mem/alloc mem/pointer-size session)
          num-floats (function-handle (mem/address-of out-floats))
          floats-addr (mem/read-address out-floats)
          floats-slice (mem/slice-global floats-addr (unchecked-multiply-int mem/float-size num-floats))]
      ;; Using a try/finally to perform an operation when the stack frame exits,
      ;; but not to try to catch anything.
      (try
        (loop [floats (transient [])
               index 0]
          (if (>= index num-floats)
            (persistent! floats)
            (recur (conj! floats (mem/read-float floats-slice
                                                 (unchecked-multiply-int index mem/float-size)
                                                 mem/big-endian))
                   (unchecked-inc-int index))))
        (finally
          (release-floats floats-addr))))))

The above code manually performs all memory operations rather than relying on coffi's dispatch. This will be more performant, but because multimethod overhead is usually relatively low, it's recommended to use the multimethod variants for convenience in colder functions.

Data Model

In addition to the macros and functions provided to build a Clojure API for native libraries, facilities are provided for taking data and loading all the symbols specified by it. This can be useful if a library provides (or an external provider maintains) a data representation of their API, as Clojure data to represent it may be programmatically generated from these sources.

The data to represent an API is a map with the following form:

(def strlen-libspec
  {:strlen {:type :function
            :symbol "strlen"
            :function/args [::mem/c-string]
            :function/ret ::mem/long}})

Each key in this map represents a single symbol to be loaded. The value is a map with at least the keys :type and :symbol. These are the currently recognized types:

function
varargs-factory
const
static-var

Each one has its own set of additional keys which can be added to the map. Both function and varargs-factory have the three keys :function/args, :function/ret, and :function/raw-fn?. The const type has :const/type and static-var has :static-var/type.

This data can be passed to the function reify-libspec, which will take the data and return a map from the same keys as the input map to whatever value is appropriate for a given symbol type (e.g. a Clojure function for function, a value for const, etc.).

(ffi/reify-libspec strlen-libspec)
;; => {:strlen #function[...]}

This functionality can be extended by specifying new types as implementations of the multimethod reify-symbolspec, although it's recommended that for any library authors who do so, namespaced keywords be used to name types.

Alternatives

This library is not the only Clojure library providing access to native code. In addition the following libraries exist:

Dtype-next has support for Java versions 8-16 and GraalVM, but is focused strongly on array-based programming, as well as being focused on keeping memory in the native side rather than marshaling data to and from Clojure-native structures. In Java 16, this uses the first iteration of Panama, while in other Java versions it uses JNA.

Tech.jna and clojure-jna both use the JNA library in all cases, and neither provide support for dealing with struct types or callbacks.

An additional alternative to coffi is to directly use the JNI, which is the longest-standing method of wrapping native code in the JVM, but comes with the downside that it requires you to write both native and Java code to use, even if you only intend to use it from Clojure.

If your application needs to be able to run in earlier versions of the JVM than 17, or you don't want to use incubator functionality, you should consider these other options. Dtype-next provides the most robust support for native code, but if you are wrapping a simple library then the other libraries may be more appealing, as they have a smaller API surface area and it's easier to wrap functions.

Benchmarks

BENCHMARKS FOR COFFI AND DTYPE-NEXT ARE BASED ON AN OLD VERSION. NEW BENCHMARKS WILL BE CREATED WHEN PANAMA COMES OUT OF PREVIEW

An additional consideration when thinking about alternatives is the performance of each available option. It's an established fact that JNA (used by all three alternative libraries on JDK <16) introduces more overhead when calling native code than JNI does.

In order to provide a benchmark to see how much of a difference the different native interfaces make, we can use criterium to benchmark each. GLFW's glfwGetTime function will be used for the test as it performs a simple operation, and is conveniently already wrapped in JNI by the excellent LWJGL library.

The following benchmarks were run on a Lenovo Thinkpad with an Intel i7-10610U running Manjaro Linux, using Clojure 1.10.3 on Java 17.

JNI

The baseline for performance is the JNI. Using LWJGL it's relatively simple to benchmark. The following Clojure CLI command will start a repl with LWJGL and criterium loaded.

$ clj -Sdeps '{:deps {org.lwjgl/lwjgl {:mvn/version "3.2.3"}
                      org.lwjgl/lwjgl-glfw {:mvn/version "3.2.3"}
                      org.lwjgl/lwjgl$natives-linux {:mvn/version "3.2.3"}
                      org.lwjgl/lwjgl-glfw$natives-linux {:mvn/version "3.2.3"}
                      criterium/criterium {:mvn/version "0.4.6"}}}'

Then from the repl

user=> (import 'org.lwjgl.glfw.GLFW)
org.lwjgl.glfw.GLFW
user=> (require '[criterium.core :as bench])
nil
user=> (GLFW/glfwInit)
true
user=> (bench/bench (GLFW/glfwGetTime) :verbose)
amd64 Linux 5.10.68-1-MANJARO 8 cpu(s)
OpenJDK 64-Bit Server VM 17+35-2724
Runtime arguments: -Dclojure.basis=/home/jsusk/.clojure/.cpcache/2667074721.basis
Evaluation count : 1613349900 in 60 samples of 26889165 calls.
      Execution time sample mean : 32.698446 ns
             Execution time mean : 32.697811 ns
Execution time sample std-deviation : 1.274600 ns
    Execution time std-deviation : 1.276437 ns
   Execution time lower quantile : 30.750813 ns ( 2.5%)
   Execution time upper quantile : 33.757662 ns (97.5%)
                   Overhead used : 6.400704 ns
nil

GLFW requires that we initialize it before calling the glfwGetTime function. Besides that this is a simple interop call which directly maps to the native function.

This gives us a basis of 32.7 ns +/-1.3 ns. All other libraries will be evaluated relative to this result.

To ensure fairness, we'll also get that overhead value to be used in further tests.

user=> bench/estimated-overhead-cache
6.400703613065185E-9

Coffi

The dependencies when using coffi are simpler, but it also requires some JVM options to support the foreign access api.

$ clj -Sdeps '{:deps {org.suskalo/coffi {:mvn/version "0.1.205"}
                      criterium/criterium {:mvn/version "0.4.6"}}}' \
      -J--add-modules=jdk.incubator.foreign \
      -J--enable-native-access=ALL-UNNAMED

In order to ensure fair comparisons, we're going to use the same overhead value on each run, so before we do the benchmark we'll set it to the observed value from last time.

user=> (require '[criterium.core :as bench])
nil
user=> (alter-var-root #'bench/estimated-overhead-cache (constantly 6.400703613065185E-9))
6.400703613065185E-9
user=> (require '[coffi.ffi :as ffi])
nil
user=> (require '[coffi.mem :as mem])
nil
user=> (ffi/load-system-library "glfw")
nil
user=> ((ffi/cfn "glfwInit" [] ::mem/int))
1
user=> (let [f (ffi/cfn "glfwGetTime" [] ::mem/double)]
         (bench/bench (f) :verbose))
amd64 Linux 5.10.68-1-MANJARO 8 cpu(s)
OpenJDK 64-Bit Server VM 17+35-2724
Runtime arguments: --add-modules=jdk.incubator.foreign --enable-native-access=ALL-UNNAMED -Dclojure.basis=/home/jsusk/.clojure/.cpcache/72793624.basis
Evaluation count : 1657995600 in 60 samples of 27633260 calls.
      Execution time sample mean : 31.382665 ns
             Execution time mean : 31.386493 ns
Execution time sample std-deviation : 1.598571 ns
    Execution time std-deviation : 1.608818 ns
   Execution time lower quantile : 29.761194 ns ( 2.5%)
   Execution time upper quantile : 33.228276 ns (97.5%)
                   Overhead used : 6.400704 ns
nil

This result is about 1.3 ns faster, and while that is less than the standard deviation of 1.6, it's quite close to it.

Clojure-JNA

Clojure-JNA uses the JNA library, which was designed to provide Java with an easy way to access native libraries, but which is known for not having the greatest performance. Since this is an older project, I'm also including the clojure dependency to ensure the correct version is used.

$ clj -Sdeps '{:deps {org.clojure/clojure {:mvn/version "1.10.3"}
                      net.n01se/clojure-jna {:mvn/version "1.0.0"}
                      criterium/criterium {:mvn/version "0.4.6"}}}'

The naive way to call the function using Clojure-JNA is to use jna/invoke.

user=> (require '[criterium.core :as bench])
nil
user=> (alter-var-root #'bench/estimated-overhead-cache (constantly 6.400703613065185E-9))
6.400703613065185E-9
user=> (require '[net.n01se.clojure-jna :as jna])
nil
user=> (jna/invoke Integer glfw/glfwInit)
1
user=> (bench/bench (jna/invoke Double glfw/glfwGetTime) :verbose)
amd64 Linux 5.10.68-1-MANJARO 8 cpu(s)
OpenJDK 64-Bit Server VM 17+35-2724
Runtime arguments: -Dclojure.basis=/home/jsusk/.clojure/.cpcache/3229486237.basis
Evaluation count : 195948720 in 60 samples of 3265812 calls.
      Execution time sample mean : 350.335614 ns
             Execution time mean : 350.373520 ns
Execution time sample std-deviation : 24.833070 ns
    Execution time std-deviation : 24.755929 ns
   Execution time lower quantile : 300.000019 ns ( 2.5%)
   Execution time upper quantile : 365.759273 ns (97.5%)
                   Overhead used : 6.400704 ns

Found 13 outliers in 60 samples (21.6667 %)
	low-severe	 12 (20.0000 %)
	low-mild	 1 (1.6667 %)
 Variance from outliers : 53.4220 % Variance is severely inflated by outliers
nil

As you can see, this method of calling functions is very bad for performance, with call overhead dominating function runtime by an order of magnitude. That said, this isn't a completely fair comparison, nor the most realistic, because this way of calling functions looks the function up on each invocation.

To adjust for this, we'll use the jna/to-fn function to give a persistent handle to the function that we can call.

user=> (let [f (jna/to-fn Double glfw/glfwGetTime)]
         (bench/bench (f) :verbose))
amd64 Linux 5.10.68-1-MANJARO 8 cpu(s)
OpenJDK 64-Bit Server VM 17+35-2724
Runtime arguments: -Dclojure.basis=/home/jsusk/.clojure/.cpcache/3229486237.basis
Evaluation count : 611095020 in 60 samples of 10184917 calls.
      Execution time sample mean : 104.623634 ns
             Execution time mean : 104.638406 ns
Execution time sample std-deviation : 7.649296 ns
    Execution time std-deviation : 7.638963 ns
   Execution time lower quantile : 92.446016 ns ( 2.5%)
   Execution time upper quantile : 110.258832 ns (97.5%)
                   Overhead used : 6.400704 ns
nil

This is much better, but is still about 3x slower than JNI, meaning the overhead from using JNA is still bigger than the function runtime.

This performance penalty is still small in the scope of longer-running functions, and so may not be a concern for your application, but it is something to be aware of.

tech.jna

The tech.jna library is similar in scope to Clojure-JNA, however was written to fit into an ecosystem of libraries meant for array-based programming for machine learning and data science.

$ clj -Sdeps '{:deps {techascent/tech.jna {:mvn/version "4.05"}
                      criterium/criterium {:mvn/version "0.4.6"}}}'

This library is also quite simple to use, the only slightly odd thing I'm doing here is to dereference the var outside the benchmark in order to ensure it's an apples-to-apples comparison. We don't want var dereference time mucking up our benchmark.

user=> (require '[criterium.core :as bench])
nil
user=> (alter-var-root #'bench/estimated-overhead-cache (constantly 6.400703613065185E-9))
6.400703613065185E-9
user=> (require '[tech.v3.jna :as jna])
nil
user=> (jna/def-jna-fn "glfw" glfwInit "initialize glfw" Integer)
#'user/glfwInit
user=> (glfwInit)
Oct 09, 2021 10:30:50 AM clojure.tools.logging$eval1122$fn__1125 invoke
INFO: Library glfw found at [:system "glfw"]
1
user=> (jna/def-jna-fn "glfw" glfwGetTime "gets the time as a double since init" Double)
#'user/glfwGetTime
user=> (let [f @#'glfwGetTime]
         (bench/bench (f) :verbose))
amd64 Linux 5.10.68-1-MANJARO 8 cpu(s)
OpenJDK 64-Bit Server VM 17+35-2724
Runtime arguments: -Dclojure.basis=/home/jsusk/.clojure/.cpcache/2910209237.basis
Evaluation count : 323281680 in 60 samples of 5388028 calls.
      Execution time sample mean : 203.976803 ns
             Execution time mean : 203.818712 ns
Execution time sample std-deviation : 14.557312 ns
    Execution time std-deviation : 14.614080 ns
   Execution time lower quantile : 179.732593 ns ( 2.5%)
   Execution time upper quantile : 213.929374 ns (97.5%)
                   Overhead used : 6.400704 ns
nil

This version is even slower than Clojure-JNA. I'm unsure where this overhead is coming from, but I'll admit that I haven't looked at their implementations very closely.

dtype-next

The library dtype-next replaced tech.jna in the toolkit of the group working on machine learning and array-based programming, and it includes support for composite data types including structs, as well as primitive functions and callbacks.

In addition, dtype-next has two different ffi backends. First is JNA, which is usable on any JDK version, and is what we'll use for the first benchmark. Second is the Java 16 version of Project Panama, which will be shown next.

In order to use the dtype-next ffi with the JNA backend, the JNA library has to be included in the dependencies.

$ clj -Sdeps '{:deps {cnuernber/dtype-next {:mvn/version "8.032"}
                      net.java.dev.jna/jna {:mvn/version "5.8.0"}
                      criterium/criterium {:mvn/version "0.4.6"}}}'

The dtype-next library also requires some more ceremony around declaring native functions. One advantage this has is that multiple symbols with the same name can be loaded from different shared libraries, but it also does increase friction when defining native wrappers.

Some easier ways to define native wrappers are provided than what is seen here, but they share some disadvantages in documentation over the core methods provided in coffi, although they are comparable to the data model provided in coffi.

user=> (require '[criterium.core :as bench])
nil
user=> (alter-var-root #'bench/estimated-overhead-cache (constantly 6.400703613065185E-9))
6.400703613065185E-9
user=> (require '[tech.v3.datatype.ffi :as dt-ffi])
nil
user=> (def fn-defs {:glfwInit {:rettype :int32} :glfwGetTime {:rettype :float64}})
#'user/fn-defs
user=> (def library-def (dt-ffi/define-library fn-defs))
#'user/library-def
user=> (def library-instance (dt-ffi/instantiate-library library-def "/usr/lib/libglfw.so"))
#'user/library-instance
user=> (def init (:glfwInit @library-instance))
#'user/init
user=> (init)
1
user=> (let [f (:glfwGetTime @library-instance)]
         (bench/bench (f) :verbose))
amd64 Linux 5.10.68-1-MANJARO 8 cpu(s)
OpenJDK 64-Bit Server VM 17+35-2724
Runtime arguments: -Dclojure.basis=/home/jsusk/.clojure/.cpcache/643862289.basis
Evaluation count : 710822100 in 60 samples of 11847035 calls.
      Execution time sample mean : 90.900112 ns
             Execution time mean : 90.919917 ns
Execution time sample std-deviation : 6.463312 ns
    Execution time std-deviation : 6.470108 ns
   Execution time lower quantile : 79.817126 ns ( 2.5%)
   Execution time upper quantile : 95.454652 ns (97.5%)
                   Overhead used : 6.400704 ns
nil

This version of JNA usage is significantly faster than either of the other JNA libraries, but is still substantially slower than using JNI or coffi.

In addition to the JNA backend, dtype-next has a Java 16-specific backend that uses an older version of Panama. This version requires similar setup to coffi in order to run.

$ clj -Sdeps '{:deps {cnuernber/dtype-next {:mvn/version "8.032"}
                      criterium/criterium {:mvn/version "0.4.6"}}}' \
      -J--add-modules=jdk.incubator.foreign \
      -J-Dforeign.restricted=permit \
      -J--add-opens=java.base/java.lang=ALL-UNNAMED \
      -J-Djava.library.path=/usr/lib/x86_64-linux-gnu

The actual code to run the benchmark is identical to the last example, but is reproduced here for completeness.

user=> (require '[criterium.core :as bench])
nil
user=> (alter-var-root #'bench/estimated-overhead-cache (constantly 6.400703613065185E-9))
6.400703613065185E-9
user=> (require '[tech.v3.datatype.ffi :as dt-ffi])
nil
user=> (def fn-defs {:glfwInit {:rettype :int32} :glfwGetTime {:rettype :float64}})
#'user/fn-defs
user=> (def library-def (dt-ffi/define-library fn-defs))
#'user/library-def
user=> (def library-instance (dt-ffi/instantiate-library library-def "/usr/lib/libglfw.so"))
#'user/library-instance
user=> (def init (:glfwInit @library-instance))
#'user/init
user=> (init)
1
user=> (let [f (:glfwGetTime @library-instance)]
         (bench/bench (f) :verbose))
amd64 Linux 5.10.68-1-MANJARO 8 cpu(s)
OpenJDK 64-Bit Server VM 16.0.2+7
Runtime arguments: --add-modules=jdk.incubator.foreign -Dforeign.restricted=permit --add-opens=java.base/java.lang=ALL-UNNAMED -Djava.library.path=/usr/lib/x86_64-linux-gnu -Dclojure.basis=/home/jsusk/.clojure/.cpcache/2337051659.basis
Evaluation count : 1588513080 in 60 samples of 26475218 calls.
      Execution time sample mean : 58.732468 ns
             Execution time mean : 58.647361 ns
Execution time sample std-deviation : 9.732389 ns
    Execution time std-deviation : 9.791738 ns
   Execution time lower quantile : 31.318115 ns ( 2.5%)
   Execution time upper quantile : 65.449222 ns (97.5%)
                   Overhead used : 6.400704 ns

Found 14 outliers in 60 samples (23.3333 %)
	low-severe	 8 (13.3333 %)
	low-mild	 4 (6.6667 %)
	high-mild	 2 (3.3333 %)
 Variance from outliers : 87.6044 % Variance is severely inflated by outliers
nil

Not reproduced here, but notable for comparison, in my testing Java 16's version of the JNI version performed about the same.

This is significantly faster than the JNA version of dtype-next, but it is still slower than modern Panama. This is likely to simply be a result of optimizations and changes to the Panama API, and when dtype-next is updated to use the Java 17 version of Panama I expect it will perform in line with coffi, but this benchmark will be reproduced when this happens. Still, this shows that as it stands, coffi is the fastest FFI available to Clojure developers.

Known Issues

The project author is aware of these issues and plans to fix them in a future release:

No known issues, hooray!

Future Plans

These features are planned for future releases.

Support for va_args type
Header parsing tool for generating a data model?
Generic type aliases
Unsigned integer types
Record-based struct types
Helper macro for out arguments
Improve error messages from defcfn macro
Mapped memory
Helper macros for custom serde implementations for composite data types

Future JDKs

The purpose of coffi is to provide a wrapper for published versions of Project Panama, starting with JDK 17. As new JDKs are released, coffi will be ported to the newer versions of Panama. Version 0.4.341 is the last version compatible with JDK 17. Version 0.5.357 is the last version compatible with JDK 18. Version 0.6.409 is the latest version compatible with JDK 19. Bugfixes, and potential backports of newer coffi features may be found on the jdk17-lts branch. Development of new features and fixes as well as support for new Panama idioms and features will continue with focus only on the latest JDK. If a particular feature is not specific to the newer JDK, PRs backporting it to versions of coffi supporting Java 17 will likely be accepted.

1.0 Release

Because the feature that coffi wraps in the JDK is in preview as of JDK 19, coffi itself will not be released in a 1.0.x version until the feature becomes a core part of the JDK, likely before or during the next LTS release, Java 21, in September 2023.

License

Distributed under the Eclipse Public License version 1.0.

IGJoshua / coffi