Using the Faust Compiler
While the Faust compiler is available in different forms (e.g., Embedded Compiler, etc.), its most "common" one is the command line version, which can be invoked using the faust command. It translates a Faust program into code in a wide range of languages (C, C++, Rust, LLVM IR, WebAssembly, etc.). The generated code can be wrapped into an optional architecture file allowing to directly produce a fully operational program.
A typical call of the Faust command line compiler is:
faust [OPTIONS] faustFile.dsp
The Faust compiler outputs C++ code by default therefore running:
faust noise.dsp
will compile noise.dsp and output the corresponding C++ code on the standard output. The option -o allows to reroute the standard output to a file:
faust noise.dsp -o noise.cpp
The -a option allows us to wrap the generated code into an architecture file:
faust -a alsa-gtk.cpp noise.dsp
which can either be placed in the same folder as the current Faust file (noise.dsp here) or be one of the standard Faust architectures.
To compile a Faust program into an ALSA application on Linux, the following commands can be used:
faust -a alsa-gtk.cpp noise.dsp -o noise.cpp
g++ -lpthread -lasound `pkg-config --cflags --libs gtk+-2.0` noise.cpp -o noise
Note that a wide range of faust2... compilation scripts can be used to facilitate this operation by taking a Faust file and returning the corresponding binary for your platform.
Structure of the Generated Code
A Faust DSP C++ class derives from the base dsp class defined as below (a similar structure is used for languages other than C++):
class dsp {
public:
dsp() {}
virtual ~dsp() {}
// Returns the number of inputs of the Faust program
virtual int getNumInputs() = 0;
// Returns the number of outputs of the Faust program
virtual int getNumOutputs() = 0;
// This method can be called to retrieve the UI description of
// the Faust program and its associated fields
virtual void buildUserInterface(UI* ui_interface) = 0;
// Returns the current sampling rate
virtual int getSampleRate() = 0;
// Init methods
virtual void init(int sample_rate) = 0;
virtual void instanceInit(int sample_rate) = 0;
virtual void instanceConstants(int sample_rate) = 0;
virtual void instanceResetUserInterface() = 0;
virtual void instanceClear() = 0;
// Returns a clone of the instance
virtual dsp* clone() = 0;
// Retrieve the global metadata of the Faust program
virtual void metadata(Meta* m) = 0;
// Compute one audio buffer
virtual void compute(int count, FAUSTFLOAT** inputs, FAUSTFLOAT** outputs) = 0;
// Compute a time-stamped audio buffer
virtual void compute(double /*date_usec*/, int count, FAUSTFLOAT** inputs,
FAUSTFLOAT** outputs) { compute(count, inputs, outputs); }
};
Methods are filled by the compiler with the actual code. In the case of noise.dsp:
class mydsp : public dsp {
private:
int iRec0[2];
int fSampleRate;
public:
void metadata(Meta* m) {
m->declare("author", "GRAME");
m->declare("filename", "noise");
m->declare("name", "Noise");
m->declare("noises.lib/name", "Faust Noise Generator Library");
m->declare("noises.lib/version", "0.0");
}
virtual int getNumInputs() {
return 0;
}
virtual int getNumOutputs() {
return 1;
}
static void classInit(int sample_rate) {}
virtual void instanceConstants(int sample_rate) {
fSampleRate = sample_rate;
}
virtual void instanceResetUserInterface() {}
virtual void instanceClear() {
for (int l0 = 0; (l0 < 2); l0 = (l0 + 1)) {
iRec0[l0] = 0;
}
}
virtual void init(int sample_rate) {
classInit(sample_rate);
instanceInit(sample_rate);
}
virtual void instanceInit(int sample_rate) {
instanceConstants(sample_rate);
instanceResetUserInterface();
instanceClear();
}
virtual mydsp* clone() {
return new mydsp();
}
virtual int getSampleRate() {
return fSampleRate;
}
virtual void buildUserInterface(UI* ui_interface) {
ui_interface->openVerticalBox("Noise");
ui_interface->closeBox();
}
virtual void compute(int count, FAUSTFLOAT** inputs, FAUSTFLOAT** outputs) {
FAUSTFLOAT* output0 = outputs[0];
for (int i = 0; (i < count); i = (i + 1)) {
iRec0[0] = ((1103515245 * iRec0[1]) + 12345);
output0[i] = FAUSTFLOAT((4.65661287e-10f * float(iRec0[0])));
iRec0[1] = iRec0[0];
}
}
};
Several fine-grained initialization methods are available:
- the
instanceInitmethod calls several additional initialization methods. - the
instanceConstantsmethod sets the instance constant state. - the
instanceClearmethod resets the instance dynamic state (delay lines...). - the
instanceResetUserInterfacemethod resets all control value to their default state.
All of those methods can be used individually on an allocated instance to reset part of its state.
The init method combines class static state and instance initialization.
When using a single instance, calling init is the simplest way to do "what is needed." When using several instances, all of them can be initialized using instanceInit, with a single call to classInit to initialize the static shared state.
The compute method takes the number of frames to process, and inputs and outputs buffers as arrays of separated mono channels. Note that by default inputs and outputs buffers are supposed to be distinct memory zones, so one cannot safely write compute(count, inputs, inputs). The -inpl compilation option can be used for that, but only in scalar mode for now.
By default the generated code process float type samples. This can be changed using the -double option (or even -quad in some backends). The FAUSTFLOAT type used in the compute method is defined in architecture files, and can be float or double, depending of the audio driver layer. Sample adaptation may have to be used between the DSP sample type and the audio driver sample type.
Controlling Code Generation
Several options of the Faust compiler allow to control the generated C++ code. By default computation is done sample by sample in a single loop. But the compiler can also generate vector and parallel code.
Vector Code Generation
Modern C++ compilers are able to do autovectorization, that is to use SIMD instructions to speedup the code. These instructions can typically operate in parallel on short vectors of 4 or 8 simple precision floating point numbers, leading to a theoretical speedup of 4 or 8.
Autovectorization of C/C++ programs is a difficult task. Current compilers are very sensitive to the way the code is arranged. In particular, complex loops can prevent autovectorization. The goal of the vector code generation is to rearrange the C++ code in a way that facilitates the autovectorization job of the C++ compiler. Instead of generating a single sample computation loop, it splits the computation into several simpler loops that communicates by vectors.
The vector code generation is activated by passing the --vectorize (or -vec) option to the Faust compiler. Two additional options are available: --vec-size <n> controls the size of the vector (by default 32 samples) and --loop-variant 0/1 gives some additional control on the loops: --loop-variant 0 generates fixed-size sub-loops with a final sub-loop that processes the last samples, --loop-variant 1 generates sub-loops of variable vector size.
To illustrate the difference between scalar code and vector code, let's take the computation of the RMS (Root Mean Square) value of a signal. Here is the Faust code that computes the Root Mean Square of a sliding window of 1000 samples:
The corresponding compute() method generated in scalar mode is the following:
virtual void compute(int count, FAUSTFLOAT** inputs, FAUSTFLOAT** outputs) {
FAUSTFLOAT* input0 = inputs[0];
FAUSTFLOAT* output0 = outputs[0];
for (int i = 0; (i < count); i = (i + 1)) {
int iTemp0 = int((1048576.0f * mydsp_faustpower2_f(float(input0[i]))));
iVec0[(IOTA & 1023)] = iTemp0;
iRec0[0] = ((iRec0[1] + iTemp0) - iVec0[((IOTA - 1000) & 1023)]);
output0[i] = FAUSTFLOAT(std::sqrt((9.53674362e-10f * float(iRec0[0]))));
IOTA = (IOTA + 1);
iRec0[1] = iRec0[0];
}
}
The -vec option leads to the following reorganization of the code:
virtual void compute(int count, FAUSTFLOAT** inputs, FAUSTFLOAT** outputs) {
fInput0_ptr = inputs[0];
FAUSTFLOAT* fInput0 = 0;
fOutput0_ptr = outputs[0];
FAUSTFLOAT* fOutput0 = 0;
int iRec0_tmp[36];
int* iRec0 = &iRec0_tmp[4];
int fullcount = count;
int index = 0;
/* Main loop */
for (index = 0; (index <= (fullcount - 32)); index = (index + 32)) {
fInput0 = &fInput0_ptr[index];
fOutput0 = &fOutput0_ptr[index];
int count = 32;
/* Vectorizable loop 0 */
/* Pre code */
iYec0_idx = ((iYec0_idx + iYec0_idx_save) & 2047);
/* Compute code */
for (int i = 0; (i < count); i = (i + 1)) {
iYec0[((i + iYec0_idx) & 2047)] =
int((1048576.0f mydsp_faustpower2_f(float(fInput0[i]))));
}
/* Post code */
iYec0_idx_save = count;
/* Recursive loop 1 */
/* Pre code */
for (int j0 = 0; (j0 < 4); j0 = (j0 + 1)) {
iRec0_tmp[j0] = iRec0_perm[j0];
}
/* Compute code */
for (int i = 0; (i < count); i = (i + 1)) {
iRec0[i] = ((iRec0[(i - 1)] + iYec0[((i + iYec0_idx) & 2047)]) -
iYec0[(((i + iYec0_idx) - 1000) & 2047)]);
}
/* Post code */
for (int j = 0; (j < 4); j = (j + 1)) {
iRec0_perm[j] = iRec0_tmp[(count + j)];
}
/* Vectorizable loop 2 */
/* Compute code */
for (int i = 0; (i < count); i = (i + 1)) {
fOutput0[i] = FAUSTFLOAT(std::sqrt((9.53674362e-10f * float(iRec0[i]))));
}
}
/* Remaining frames */
if (index < fullcount) {
fInput0 = &fInput0_ptr[index];
fOutput0 = &fOutput0_ptr[index];
int count = (fullcount - index);
/* Vectorizable loop 0 */
/* Pre code */
iYec0_idx = ((iYec0_idx + iYec0_idx_save) & 2047);
/* Compute code */
for (int i = 0; (i < count); i = (i + 1)) {
iYec0[((i + iYec0_idx) & 2047)] = int((1048576.0f *
mydsp_faustpower2_f(float(fInput0[i]))));
}
/* Post code */
iYec0_idx_save = count;
/* Recursive loop 1 */
/* Pre code */
for (int j0 = 0; (j0 < 4); j0 = (j0 + 1)) {
iRec0_tmp[j0] = iRec0_perm[j0];
}
/* Compute code */
for (int i = 0; (i < count); i = (i + 1)) {
iRec0[i] = ((iRec0[(i - 1)] + iYec0[((i + iYec0_idx) & 2047)]) -
iYec0[(((i + iYec0_idx) - 1000) & 2047)]);
}
/* Post code */
for (int j = 0; (j < 4); j = (j + 1)) {
iRec0_perm[j] = iRec0_tmp[(count + j)];
}
/* Vectorizable loop 2 */
/* Compute code */
for (int i = 0; (i < count); i = (i + 1)) {
fOutput0[i] = FAUSTFLOAT(std::sqrt((9.53674362e-10f * float(iRec0[i]))));
}
}
}
While the second version of the code is more complex, it turns out to be much easier to vectorize efficiently by the C++ compiler. With the exact same compilation options: -O3 -xHost -ftz -fno-alias -fp-model fast=2, the scalar version leads to a throughput performance of 129.144 MB/s, while the vector version achieves 359.548 MB/s, a speedup of x2.8 !
The vector code generation is built on top of the scalar code generation (see previous figure). Every time an expression needs to be compiled, the compiler checks if it requires a separate loop or not. Expressions that are shared (and are complex enough) are good candidates to be compiled in a separate loop, as well as recursive expressions and expressions used in delay lines.
The result is a directed graph in which each node is a computation loop (see figure below). This graph is stored in the class object and a topological sort is applied to it before printing the code.
Parallel Code Generation
Parallel code generation is activated by passing either the --openMP (or -omp) option or the --scheduler (or -sch) option. It implies that the -vec option as well as the parallel code generation are built on top of the vector code generation.
The OpenMP Code Generator
The --openMP (or -omp) option, when given to the Faust compiler, will insert appropriate OpenMP directives into the C++ code. OpenMP is a well established API that is used to explicitly define direct multi-threaded, shared memory parallelism. It is based on a fork-join model of parallelism (see figure above). Parallel regions are delimited by #pragma omp parallel constructs. At the entrance of a parallel region, a group of parallel threads is activated. The code within a parallel region is executed by each thread of the parallel group until the end of the region.
#pragma omp parallel
{
// the code here is executed simultaneously by every thread of the parallel
// team
...
}
In order not to have every thread doing redundantly the exact same work, OpenMP provides specific work-sharing directives. For example #pragma omp sections allows to break the work into separate, discrete sections, each section being executed by one thread:
#pragma omp parallel
{
#pragma omp sections
{
#pragma omp section
{
// job 1
}
#pragma omp section
{
// job 2
}
...
}
...
}
Adding Open MP Directives
As said before, parallel code generation is built on top of vector code generation. The graph of loops produced by the vector code generator is topologically sorted in order to detect the loops that can be computed in parallel. The first set S_0 (loops L1, L2 and L3) contains the loops that don't depend on any other loops, the set S_1 contains the loops that only depend on loops of S_0, (that is loops L4 and L5), etc..
As all the loops of a given set S_n can be computed in parallel, the compiler will generate a sections construct with a section for each loop.
#pragma omp sections
{
#pragma omp section
for (...) {
// Loop 1
}
#pragma omp section
for (...) {
// Loop 2
}
...
}
If a given set contains only one loop, then the compiler checks to see if the loop can be parallelized (no recursive dependencies) or not. If it can be parallelized, it generates:
#pragma omp for
for (...) {
// Loop code
}
otherwise it generates a single construct so that only one thread will execute the loop:
#pragma omp single
for (...) {
// Loop code
}
Example of Parallel OpenMP Code
To illustrate how Faust uses the OpenMP directives, here is a very simple example, two 1-pole filters in parallel connected to an adder:
The corresponding compute() method obtained using the -omp option looks like this:
virtual void compute(int fullcount, FAUSTFLOAT** inputs, FAUSTFLOAT** outputs) {
float fRec0_tmp[36];
float fRec1_tmp[36];
FAUSTFLOAT* fInput0 = 0;
FAUSTFLOAT* fInput1 = 0;
FAUSTFLOAT* fOutput0 = 0;
float* fRec0 = &fRec0_tmp[4];
float* fRec1 = &fRec1_tmp[4];
fInput0_ptr = inputs[0];
fInput1_ptr = inputs[1];
fOutput0_ptr = outputs[0];
#pragma omp parallel\
firstprivate(fInput0, fInput1, fOutput0, fRec0, fRec1)
{
for (int index = 0; (index < fullcount); index = (index + 32)) {
fInput0 = &fInput0_ptr[index];
fInput1 = &fInput1_ptr[index];
fOutput0 = &fOutput0_ptr[index];
int count = min(32, (fullcount - index));
#pragma omp sections
{
#pragma omp section
{
/* Recursive loop 0 */
/* Pre code */
for (int j0 = 0; (j0 < 4); j0 = (j0 + 1)) {
fRec0_tmp[j0] = fRec0_perm[j0];
}
/* Compute code */
for (int i = 0; (i < count); i = (i + 1)) {
fRec0[i] = ((0.899999976f * fRec0[(i - 1)]) +
(0.100000001f * float(fInput0[i])));
}
/* Post code */
for (int j = 0; (j < 4); j = (j + 1)) {
fRec0_perm[j] = fRec0_tmp[(count + j)];
}
}
#pragma omp section
{
/* Recursive loop 1 */
/* Pre code */
for (int j1 = 0; (j1 < 4); j1 = (j1 + 1)) {
fRec1_tmp[j1] = fRec1_perm[j1];
}
/* Compute code */
for (int i = 0; (i < count); i = (i + 1)) {
fRec1[i] = ((0.899999976f * fRec1[(i - 1)]) +
(0.100000001f * float(fInput1[i])));
}
/* Post code */
for (int j = 0; (j < 4); j = (j + 1)) {
fRec1_perm[j] = fRec1_tmp[(count + j)];
}
}
}
#pragma omp single
{
/* Vectorizable loop 2 */
/* Compute code */
for (int i = 0; (i < count); i = (i + 1)) {
fOutput0[i] = FAUSTFLOAT((fRec0[i] + fRec1[i]));
}
}
}
}
}
This code requires some comments:
- the parallel construct
#pragma omp parallelis the fundamental construct that starts parallel execution. The number of parallel threads is generally the number of CPU cores but it can be controlled in several ways. - variables external to the parallel region are shared by default. The pragma
firstprivate(fRec0,fRec1)indicates that each thread should have its private copy offRec0andfRec1. The reason is that accessing shared variables requires an indirection and is quite inefficient compared to private copies. - the top level loop
for (int index = 0;...)...is executed by all threads simultaneously. The subsequent work-sharing directives inside the loop will indicate how the work must be shared between threads. - please note that an implied barrier exists at the end of each work-sharing region. All threads must have executed the barrier before any of them can continue.
- the work-sharing directive
#pragma omp singleindicates that this first section will be executed by only one thread (any of them). - the work-sharing directive
#pragma omp sectionsindicates that each corresponding#pragma omp section, here our two filters, will be executed in parallel. - the loop construct
#pragma omp forspecifies that the iterations of the associated loop will be executed in parallel. The iterations of the loop are distributed across the parallel threads. For example, if we have two threads, the first one can compute indices between 0 and count/2 and the other one between count/2 and count. - finally
#pragma omp singleindicates that this section will be executed by only one thread (any of them).
The Scheduler Code Generator
With the --scheduler (or -sch) option given to the Faust compiler, the computation graph is cut into separate computation loops (called "tasks"), and a "Work Stealing Scheduler" is used to activate and execute them following their dependencies. A pool of worked threads is created and each thread uses it's own local WSQ (Work Stealing Queue) of tasks. A WSQ is a special queue with a Push operation, a "private" LIFO Pop operation and a "public" FIFO Pop operation.
Starting from a ready task, each thread follows the dependencies, possibly pushing ready sub-tasks into it's own local WSQ. When no more tasks can be activated on a given computation path, the thread pops a task from it's local WSQ. If the WSQ is empty, then the thread is allowed to "steal" tasks from other threads WSQ.
The local LIFO Pop operation allows better cache locality and the FIFO steal Pop "larger chuck" of work to be done. The reason for this is that many work stealing workloads are divide-and-conquer in nature, stealing one of the oldest task implicitly also steals a (potentially) large sub-tree of computations that will unfold once that piece of work is stolen and run.
Compared to the OpenMP model (-omp) the new model is worse for simple Faust programs and usually starts to behave comparable or sometimes better for "complex enough" Faust programs. In any case, since OpenMP does not behave so well with GCC compilers, and is unusable on OSX in real-time contexts, this new scheduler option has it's own value. We plan to improve it adding a "pipelining" idea in the future.
Example of Parallel Scheduler Code
To illustrate how Faust generates the scheduler code, let's reuse the previous example made of two 1-pole filters in parallel connected to an adder:
When -sch option is used, the content of the additional architecture/scheduler.h file is inserted in the generated code. It contains code to deal with WSQ and thread management. The compute() and computeThread() methods are the following:
virtual void compute(int count, FAUSTFLOAT** inputs, FAUSTFLOAT** outputs) {
fInput0_ptr = inputs[0];
fInput1_ptr = inputs[1];
fOutput0_ptr = outputs[0];
fCount = count;
fIndex = 0;
/* End task has only one input, so will be directly activated */
/* Only initialize tasks with more than one input */
initTask(fScheduler, 4, 2);
/* Push ready tasks in each thread WSQ */
initTaskList(fScheduler, -1);
signalAll(fScheduler);
computeThread(0);
syncAll(fScheduler);
}
void computeThread(int num_thread) {
int count = fCount;
FAUSTFLOAT* fInput0 = 0;
FAUSTFLOAT* fInput1 = 0;
FAUSTFLOAT* fOutput0 = 0;
int tasknum = 0;
while ((fIndex < fCount)) {
fInput0 = &fInput0_ptr[fIndex];
fInput1 = &fInput1_ptr[fIndex];
fOutput0 = &fOutput0_ptr[fIndex];
count = min(32, (fCount - fIndex));
switch (tasknum) {
case 0: {
/* Work Stealing task */
tasknum = getNextTask(fScheduler, num_thread);
break;
}
case 1: {
/* Last task */
fIndex = (fIndex + 32);
if (fIndex < fCount) {
/* End task has only one input, so will be directly activated */
/* Only initialize tasks with more than one input */
initTask(fScheduler, 4, 2);
/* Push ready tasks in 'num_thread' WSQ */
initTaskList(fScheduler, num_thread);
}
tasknum = 0;
break;
}
case 2: {
/* Recursive loop 2 */
/* Pre code */
for (int j0 = 0; (j0 < 4); j0 = (j0 + 1)) {
fRec0_tmp[j0] = fRec0_perm[j0];
}
/* Compute code */
for (int i = 0; (i < count); i = (i + 1)) {
fRec0[i] = ((0.899999976f * fRec0[(i - 1)]) +
(0.100000001f * float(fInput0[i])));
}
/* Post code */
for (int j = 0; (j < 4); j = (j + 1)) {
fRec0_perm[j] = fRec0_tmp[(count + j)];
}
/* One output only */
activateOneOutputTask(fScheduler, num_thread, 4, &tasknum);
break;
}
case 3: {
/* Recursive loop 3 */
/* Pre code */
for (int j1 = 0; (j1 < 4); j1 = (j1 + 1)) {
fRec1_tmp[j1] = fRec1_perm[j1];
}
/* Compute code */
for (int i = 0; (i < count); i = (i + 1)) {
fRec1[i] = ((0.899999976f * fRec1[(i - 1)]) +
(0.100000001f * float(fInput1[i])));
}
/* Post code */
for (int j = 0; (j < 4); j = (j + 1)) {
fRec1_perm[j] = fRec1_tmp[(count + j)];
}
/* One output only */
activateOneOutputTask(fScheduler, num_thread, 4, &tasknum);
break;
}
case 4: {
/* Vectorizable loop 4 */
/* Compute code */
for (int i = 0; (i < count); i = (i + 1)) {
fOutput0[i] = FAUSTFLOAT((fRec0[i] + fRec1[i]));
}
tasknum = 1;
break;
}
}
}
}