diff options
author | Ian Bertolacci <ian.bertolacci@gmail.com> | 2015-08-02 15:37:01 -0700 |
---|---|---|
committer | Ian Bertolacci <ian.bertolacci@gmail.com> | 2015-08-02 15:37:01 -0700 |
commit | b27d5268227f8812c894a217a08fd8c6221d1797 (patch) | |
tree | 6e2431382e42df356cc2c82383c4a0a98de465b2 | |
parent | 1f6f7b784345a7224fef64bc27db206a0946b366 (diff) |
Added modules and main()
Worth noting that this will change the diff of the parallel section quite a bit, since they became the body of the main procedure.
Thus each line in intented.
-rw-r--r-- | chapel.html.markdown | 462 |
1 files changed, 258 insertions, 204 deletions
diff --git a/chapel.html.markdown b/chapel.html.markdown index cfde0454..d8783728 100644 --- a/chapel.html.markdown +++ b/chapel.html.markdown @@ -673,234 +673,288 @@ var copyNewTypeList = new GenericClass( realList, int ); for value in copyNewTypeList do write( value, ", " ); writeln( ); -// Parallelism -// In other languages, parallelism is typically this is done with -// complicated libraries and strange class structure hierarchies. -// Chapel has it baked right into the language. - -// A begin statement will spin the body of that statement off into one new task. -// A sync statement will ensure that the progress of the main -// task will not progress until the children have synced back up. -sync { - begin { // Start of new task's body - var a = 0; - for i in 1..1000 do a += 1; - writeln( "Done: ", a); - } // End of new tasks body - writeln( "spun off a task!"); -} -writeln( "Back together" ); -proc printFibb( n: int ){ - writeln( "fibonacci(",n,") = ", fibonacci( n ) ); -} +// Modules are Chapel's way of managing name spaces. +// The files containing these modules do not need to be named after the modules +// (as is with Java), but files implicitly name modules. +// In this case, this file implicitly names the 'learnchapel' module -// A cobegin statement will spin each statement of the body into one new task -cobegin { - printFibb( 20 ); // new task - printFibb( 10 ); // new task - printFibb( 5 ); // new task - { - // This is a nested statement body and thus is a single statement - // to the parent statement and is executed by a single task - writeln( "this gets" ); - writeln( "executed as" ); - writeln( "a whole" ); +module OurModule { + // We can use modules inside of other modules. + use Time; + + // We'll use this a procedure in the parallelism section. + proc countdown( seconds: int ){ + for i in 1..seconds by -1 { + writeln( i ); + sleep( 1 ); + } + } + + // Submodule of Ourmodule + // It is possible to create arbitrarily deep module nests. + module ChildModule { + proc foo(){ + writeln( "ChildModule.foo()"); + } } -} -// Notice here that the prints from each statement may happen in any order. - -// Coforall loop will create a new task for EACH iteration -var num_tasks = 10; // Number of tasks we want -coforall taskID in 1..#num_tasks { - writeln( "Hello from task# ", taskID ); -} -// Again we see that prints happen in any order. -// NOTE! coforall should be used only for creating tasks! -// Using it to iterating over a structure is very a bad idea! - -// forall loops are another parallel loop, but only create a smaller number -// of tasks, specifically --dataParTasksPerLocale=number of task -forall i in 1..100 { - write( i, ", "); -} -writeln( ); -// Here we see that there are sections that are in order, followed by -// a section that would not follow ( e.g. 1, 2, 3, 7, 8, 9, 4, 5, 6, ). -// This is because each task is taking on a chunk of the range 1..10 -// (1..3, 4..6, or 7..9) doing that chunk serially, but each task happens -// in parallel. -// Your results may depend on your machine and configuration - -// For both the forall and coforall loops, the execution of the -// parent task will not continue until all the children sync up. - -// forall loops are particularly useful for parallel iteration over arrays. -// Lets run an experiment to see how much faster a parallel loop is -use Time; // Import the Time module to use Timer objects -var timer: Timer; -var myBigArray: [{1..4000,1..4000}] real; // Large array we will write into - -// Serial Experiment -timer.start( ); // Start timer -for (x,y) in myBigArray.domain { // Serial iteration - myBigArray[x,y] = (x:real) / (y:real); -} -timer.stop( ); // Stop timer -writeln( "Serial: ", timer.elapsed( ) ); // Print elapsed time -timer.clear( ); // Clear timer for parallel loop - -// Parallel Experiment -timer.start( ); // start timer -forall (x,y) in myBigArray.domain { // Parallel iteration - myBigArray[x,y] = (x:real) / (y:real); -} -timer.stop( ); // Stop timer -writeln( "Parallel: ", timer.elapsed( ) ); // Print elapsed time -timer.clear( ); -// You may have noticed that (depending on how many cores you have) -// that the parallel loop went faster than the serial loop - -// A succinct way of writing a forall loop over an array: -// iterate over values -[ val in myBigArray ] val = 1 / val; - -// or iterate over indicies -[ idx in myBigArray.domain ] myBigArray[idx] = -myBigArray[idx]; - -proc countdown( seconds: int ){ - for i in 1..seconds by -1 { - writeln( i ); - sleep( 1 ); + + module SiblingModule { + proc foo(){ + writeln( "SiblingModule.foo()" ); + } } -} +} // end OurModule + +// Using OurModule also uses all the modules it uses. +// Since OurModule uses Time, we also use time. +use OurModule; + +// At this point we have not used ChildModule or SiblingModule so their symbols +// (i.e. foo ) are not available to us. +// However, the module names are, and we can explicitly call foo() through them. +SiblingModule.foo(); // Calls SiblingModule.foo() + +// Super explicit naming. +OurModule.ChildModule.foo(); // Calls ChildModule.foo() + +use ChildModule; +foo(); // Less explicit call on ChildModule.foo() + +// We can declare a main procedure +// Note: all the code above main still gets executed. +proc main(){ + + // Parallelism + // In other languages, parallelism is typically this is done with + // complicated libraries and strange class structure hierarchies. + // Chapel has it baked right into the language. + + // A begin statement will spin the body of that statement off + // into one new task. + // A sync statement will ensure that the progress of the main + // task will not progress until the children have synced back up. + sync { + begin { // Start of new task's body + var a = 0; + for i in 1..1000 do a += 1; + writeln( "Done: ", a); + } // End of new tasks body + writeln( "spun off a task!"); + } + writeln( "Back together" ); -// Atomic variables, common to many languages, are ones whose operations -// occur uninterupted. Multiple threads can both modify atomic variables -// and can know that their values are safe. -// Chapel atomic variables can be of type bool, int, uint, and real. -var uranium: atomic int; -uranium.write( 238 ); // atomically write a variable -writeln( uranium.read() ); // atomically read a variable - -// operations are described as functions, you could define your own operators. -uranium.sub( 3 ); // atomically subtract a variable -writeln( uranium.read() ); - -var replaceWith = 239; -var was = uranium.exchange( replaceWith ); -writeln( "uranium was ", was, " but is now ", replaceWith ); - -var isEqualTo = 235; -if uranium.compareExchange( isEqualTo, replaceWith ) { - writeln( "uranium was equal to ", isEqualTo, - " so replaced value with ", replaceWith ); -} else { - writeln( "uranium was not equal to ", isEqualTo, - " so value stays the same... whatever it was" ); -} + proc printFibb( n: int ){ + writeln( "fibonacci(",n,") = ", fibonacci( n ) ); + } -sync { - begin { // Reader task - writeln( "Reader: waiting for uranium to be ", isEqualTo ); - uranium.waitFor( isEqualTo ); - writeln( "Reader: uranium was set (by someone) to ", isEqualTo ); + // A cobegin statement will spin each statement of the body into one new task + cobegin { + printFibb( 20 ); // new task + printFibb( 10 ); // new task + printFibb( 5 ); // new task + { + // This is a nested statement body and thus is a single statement + // to the parent statement and is executed by a single task + writeln( "this gets" ); + writeln( "executed as" ); + writeln( "a whole" ); + } } + // Notice here that the prints from each statement may happen in any order. - begin { // Writer task - writeln( "Writer: will set uranium to the value ", isEqualTo, " in..." ); - countdown( 3 ); - uranium.write( isEqualTo ); + // Coforall loop will create a new task for EACH iteration + var num_tasks = 10; // Number of tasks we want + coforall taskID in 1..#num_tasks { + writeln( "Hello from task# ", taskID ); + } + // Again we see that prints happen in any order. + // NOTE! coforall should be used only for creating tasks! + // Using it to iterating over a structure is very a bad idea! + + // forall loops are another parallel loop, but only create a smaller number + // of tasks, specifically --dataParTasksPerLocale=number of task + forall i in 1..100 { + write( i, ", "); + } + writeln( ); + // Here we see that there are sections that are in order, followed by + // a section that would not follow ( e.g. 1, 2, 3, 7, 8, 9, 4, 5, 6, ). + // This is because each task is taking on a chunk of the range 1..10 + // (1..3, 4..6, or 7..9) doing that chunk serially, but each task happens + // in parallel. + // Your results may depend on your machine and configuration + + // For both the forall and coforall loops, the execution of the + // parent task will not continue until all the children sync up. + + // forall loops are particularly useful for parallel iteration over arrays. + // Lets run an experiment to see how much faster a parallel loop is + use Time; // Import the Time module to use Timer objects + var timer: Timer; + var myBigArray: [{1..4000,1..4000}] real; // Large array we will write into + + // Serial Experiment + timer.start( ); // Start timer + for (x,y) in myBigArray.domain { // Serial iteration + myBigArray[x,y] = (x:real) / (y:real); + } + timer.stop( ); // Stop timer + writeln( "Serial: ", timer.elapsed( ) ); // Print elapsed time + timer.clear( ); // Clear timer for parallel loop + + // Parallel Experiment + timer.start( ); // start timer + forall (x,y) in myBigArray.domain { // Parallel iteration + myBigArray[x,y] = (x:real) / (y:real); + } + timer.stop( ); // Stop timer + writeln( "Parallel: ", timer.elapsed( ) ); // Print elapsed time + timer.clear( ); + // You may have noticed that (depending on how many cores you have) + // that the parallel loop went faster than the serial loop + + // A succinct way of writing a forall loop over an array: + // iterate over values + [ val in myBigArray ] val = 1 / val; + + // or iterate over indicies + [ idx in myBigArray.domain ] myBigArray[idx] = -myBigArray[idx]; + + proc countdown( seconds: int ){ + for i in 1..seconds by -1 { + writeln( i ); + sleep( 1 ); + } } -} -// sync vars have two states: empty and full. -// If you read an empty variable or write a full variable, you are waited -// until the variable is full or empty again -var someSyncVar$: sync int; // varName$ is a convention not a law. -sync { - begin { // Reader task - writeln( "Reader: waiting to read." ); - var read_sync = someSyncVar$; - writeln( "value is ", read_sync ); + // Atomic variables, common to many languages, are ones whose operations + // occur uninterupted. Multiple threads can both modify atomic variables + // and can know that their values are safe. + // Chapel atomic variables can be of type bool, int, uint, and real. + var uranium: atomic int; + uranium.write( 238 ); // atomically write a variable + writeln( uranium.read() ); // atomically read a variable + + // operations are described as functions, you could define your own operators. + uranium.sub( 3 ); // atomically subtract a variable + writeln( uranium.read() ); + + var replaceWith = 239; + var was = uranium.exchange( replaceWith ); + writeln( "uranium was ", was, " but is now ", replaceWith ); + + var isEqualTo = 235; + if uranium.compareExchange( isEqualTo, replaceWith ) { + writeln( "uranium was equal to ", isEqualTo, + " so replaced value with ", replaceWith ); + } else { + writeln( "uranium was not equal to ", isEqualTo, + " so value stays the same... whatever it was" ); } - begin { // Writer task - writeln( "Writer: will write in..." ); - countdown( 3 ); - someSyncVar$ = 123; + sync { + begin { // Reader task + writeln( "Reader: waiting for uranium to be ", isEqualTo ); + uranium.waitFor( isEqualTo ); + writeln( "Reader: uranium was set (by someone) to ", isEqualTo ); + } + + begin { // Writer task + writeln( "Writer: will set uranium to the value ", isEqualTo, " in..." ); + countdown( 3 ); + uranium.write( isEqualTo ); + } } -} -// single vars can only be written once. A read on an unwritten single results -// in a wait, but when the variable has a value it can be read indefinitely -var someSingleVar$: single int; // varName$ is a convention not a law. -sync { - begin { // Reader task - writeln( "Reader: waiting to read." ); - for i in 1..5 { - var read_single = someSingleVar$; - writeln( "Reader: iteration ", i,", and the value is ", read_single ); + // sync vars have two states: empty and full. + // If you read an empty variable or write a full variable, you are waited + // until the variable is full or empty again + var someSyncVar$: sync int; // varName$ is a convention not a law. + sync { + begin { // Reader task + writeln( "Reader: waiting to read." ); + var read_sync = someSyncVar$; + writeln( "value is ", read_sync ); + } + + begin { // Writer task + writeln( "Writer: will write in..." ); + countdown( 3 ); + someSyncVar$ = 123; } } - begin { // Writer task - writeln( "Writer: will write in..." ); - countdown( 3 ); - someSingleVar$ = 5; // first and only write ever. + // single vars can only be written once. A read on an unwritten single results + // in a wait, but when the variable has a value it can be read indefinitely + var someSingleVar$: single int; // varName$ is a convention not a law. + sync { + begin { // Reader task + writeln( "Reader: waiting to read." ); + for i in 1..5 { + var read_single = someSingleVar$; + writeln( "Reader: iteration ", i,", and the value is ", read_single ); + } + } + + begin { // Writer task + writeln( "Writer: will write in..." ); + countdown( 3 ); + someSingleVar$ = 5; // first and only write ever. + } } -} -// Heres an example of using atomics and a synch variable to create a -// count-down mutex (also known as a multiplexer) -var count: atomic int; // our counter -var lock$: sync bool; // the mutex lock - -count.write( 2 ); // Only let two tasks in at a time. -lock$.writeXF( true ); // Set lock$ to full (unlocked) -// Note: The value doesnt actually matter, just the state -// (full:unlocked / empty:locked) -// Also, writeXF() fills (F) the sync var regardless of its state (X) - -coforall task in 1..#5 { // Generate tasks - // Create a barrier - do{ - lock$; // Read lock$ (wait) - }while count.read() < 1; // Keep waiting until a spot opens up + // Heres an example of using atomics and a synch variable to create a + // count-down mutex (also known as a multiplexer) + var count: atomic int; // our counter + var lock$: sync bool; // the mutex lock + + count.write( 2 ); // Only let two tasks in at a time. + lock$.writeXF( true ); // Set lock$ to full (unlocked) + // Note: The value doesnt actually matter, just the state + // (full:unlocked / empty:locked) + // Also, writeXF() fills (F) the sync var regardless of its state (X) + + coforall task in 1..#5 { // Generate tasks + // Create a barrier + do{ + lock$; // Read lock$ (wait) + }while count.read() < 1; // Keep waiting until a spot opens up - count.sub(1); // decrement the counter - lock$.writeXF( true ); // Set lock$ to full (signal) + count.sub(1); // decrement the counter + lock$.writeXF( true ); // Set lock$ to full (signal) - // Actual 'work' - writeln( "Task #", task, " doing work." ); - sleep( 2 ); + // Actual 'work' + writeln( "Task #", task, " doing work." ); + sleep( 2 ); - count.add( 1 ); // Increment the counter - lock$.writeXF( true ); // Set lock$ to full (signal) -} + count.add( 1 ); // Increment the counter + lock$.writeXF( true ); // Set lock$ to full (signal) + } -// we can define the operations + * & | ^ && || min max minloc maxloc -// over an entire array using scans and reductions -// Reductions apply the operation over the entire array and -// result in a single value -var listOfValues: [1..10] int = [15,57,354,36,45,15,456,8,678,2]; -var sumOfValues = + reduce listOfValues; -var maxValue = max reduce listOfValues; // 'max' give just max value - -// 'maxloc' gives max value and index of the max value -// Note: We have to zip the array and domain together with the zip iterator -var (theMaxValue, idxOfMax) = maxloc reduce zip(listOfValues, - listOfValues.domain); + // we can define the operations + * & | ^ && || min max minloc maxloc + // over an entire array using scans and reductions + // Reductions apply the operation over the entire array and + // result in a single value + var listOfValues: [1..10] int = [15,57,354,36,45,15,456,8,678,2]; + var sumOfValues = + reduce listOfValues; + var maxValue = max reduce listOfValues; // 'max' give just max value + + // 'maxloc' gives max value and index of the max value + // Note: We have to zip the array and domain together with the zip iterator + var (theMaxValue, idxOfMax) = maxloc reduce zip(listOfValues, + listOfValues.domain); -writeln( (sumOfValues, maxValue, idxOfMax, listOfValues[ idxOfMax ] ) ); - -// Scans apply the operation incrementally and return an array of the -// value of the operation at that index as it progressed through the -// array from array.domain.low to array.domain.high -var runningSumOfValues = + scan listOfValues; -var maxScan = max scan listOfValues; -writeln( runningSumOfValues ); -writeln( maxScan ); + writeln( (sumOfValues, maxValue, idxOfMax, listOfValues[ idxOfMax ] ) ); + + // Scans apply the operation incrementally and return an array of the + // value of the operation at that index as it progressed through the + // array from array.domain.low to array.domain.high + var runningSumOfValues = + scan listOfValues; + var maxScan = max scan listOfValues; + writeln( runningSumOfValues ); + writeln( maxScan ); +} ``` Who is this tutorial for? |