Questions about Chapel

• What's the minimal subset of the language capable of representing all chapel programs?
• What subset of the language will we concern ourselves with?

Questions for the Chapel Dev Team

• If a given thread for a forall loop is assigned iterations 1, 3 and 7, will they necessarily be assigned in that order?
• How do/will parallel iterators affect the assignment of threads to iterations of a forall loop.

Chapel Constructs

Concepts

• Tasks

Types

• int
• bool
• range
• rectangular domain
• array
• sync

Have hidden state, empty or full. Read blocks on empty, write blocks on full (Has other methods, writeXX(), readXX() etc.)

• single

Write once, reads block until written to (Error condition if written to more than once)

Statements

• sync <statement>

Wait for all tasks created in the statement to complete

• serial <expression> <statement>

If expression is true, serializes all code in statement

• atomic <statement>

A transaction

• for
• forall

Like coforall, but iterations distributed among tasks in any way, including, possibly, one task

• coforall

Assign one task to each iteration, wait for all iterations to complete

• begin

Launch task to execute statement, proceed without blocking

• cobegin

Assign one task to each statement, wait for all to complete

• call
• assignment
• if
• block
• select
• expression
• (do)while
• return
• procedure declaration
• variable declaration
• yield

Expressions

• Literal
• Variable
• Function call
• L-Value
• Parenthesis
• Unary operators
  • - (Negation)
• Binary Operators
  • + (Addition)
  • - (Subtraction)
  • / (Integer/Rational Divide)
  • * (Times)
  • % (Modulus)
  • | | (Logical Or)
  • & & (Logical and)
• = (Assignment)

Iterators and for/forall

The use of iterators in for/forall loops can be implemented using a manager-worker strategy. When an iterator is encountered, the caller ("manager") spawns a new process to run the iterator. In addition to any declared parameters, the iterator has an extra parameter for the process ID of the manager. Whenever the iterator encounters a yield statement, it sends the yielded value to the manager. The manager then non-deterministically distributes the yielded value to one of n worker threads, which sets the loop variable to that value and runs the loop. When the end of the iterator is reached, it sends a special termination message. The manager then sends termination messages to all works and waits for them to complete.