Skip to content

DSLX Example: Prefix Scan Computation

In this document we explain in detail the implementation of a 8 byte prefix scan computation. In order to understand the implementation, it is useful to understand the intended functionality first.

For a given input of 8 bytes, the scan iterates from left to right over the input and produces an output of the same size. Each element in the output contains the count of duplicate values seen so far in the input. The counter resets to 0 if a new value is found.

For example, for this input:

  let input = bits[8,32]:[0, 0, 0, 0, 0, 0, 0, 0]

the code should produce this output:

  bits[8,3]:[0, 1, 2, 3, 4, 5, 6, 7])

At index 0 it has not yet found any value, so it assigns a counter value of 0.

At index 1 it finds the second occurrence of the value '0' (which is the 1st duplicate) and therefore adds a 1 to the counter from index 0.

At index 2 it finds the third occurrence of the value '0' (which is the 2nd duplicate) and therefore adds a 1 to the counter from index 1. And so on.

Correspondingly, for this input:

  let input = bits[8,32]:[0, 0, 1, 1, 2, 2, 3, 3]

it should produce:

  assert_eq(result, bits[8,3]:[0, 1, 0, 1, 0, 1, 0, 1])

The full listing is in examples/dslx_intro/prefix_scan_equality.x.

Function prefix_scan_eq

The implementation displays a few interesting language features.

The function prototype is straigt-forward. Input is an array of 8 values of type u32. Output is an array of size 8 holding 3-bit values (the maximum resulting count can only be 7, which fits in 3 bits).

fn prefix_scan_eq(x: u32[8]) -> bits[8,3] {

The first let expression produces a tuple of 3 values. It only cares about the last value result, so it stubs out the other two elements via the 'ignore' placeholder _.

  let (_, _, result) =

Why a 3-Tuple? Because he following loop has tuple of three values as the accumulator. The return type of the loop is the type of the accumulator, so above let needs to be of the same type.

Enumerated Loop

Using tuples as the accumulator is a convenient way to model multiple loop-carried values:

    for ((i, elem), (prior, count, result)): ((u32, u32), (u32, u3, bits[8,3]))
          in enumerate(x) {

The iterable of this loop is enumerate(x). On each iteration, this construct delivers a tuple consisting of current index and current element. This is represented as the tuple (i, elem) in the for construct.

The loop next specifies the accumulator, which is a 3-tuple consisting of the values named prior, count, and result.

The types of the iterable and accumulator are specified next. The iterable is a tuple consisting of two u32 values. The accumulator is more interesting, it is a tuple consiting of a u32 value (prior), a u3 value (count), and a 2-dimension array type bits[8, 3], which is an array holding 8 elements of bit-width 3. This is the type of result in the accumulator.

Looping back to the prior let statement, it ignores the prior and count members of the tuple and will only return the result part.

A Match Expression

The next expression is an interesting match expression. The let expression binds the tuple (to_place, new_count): (u3, u3) to the result of the following match expression:

let (to_place, new_count): (u3, u3) = match (i == u32:0, prior == elem) {

to_place will hold the value that is to be written at a given index. new_count will contain the updated counter value.

The match expression evaluates two conditions in parallel:

  • is i == 0?
  • is the prior element the same as the current elem

Two tests mean there are four possible cases, which are all handled in the following four cases:

      // i == 0 (no matter whether prior == elem or not):
      //    we set position 0 to 0 and update the new_counter to 1
      (true, true) => (u3:0, u3:1),
      (true, false) => (u3:0, u3:1),

      // if i != 0 - if the current element is the same as pior,
      //    set to_place to the value of the current count
      //    update new_counter with the increased counter value
      (false, true) => (count, count + u3:1),

      // if i != 0 - if current element is different from prior,
      //     set to_place back to 0
      //     set new_counter back to 1
      (false, false) => (u3:0, u3:1),
    };

To update the result, we set index i in the result array to the value to_place via the built-in update function, which produces a new value new_result):

    let new_result: bits[8,3] = update(result, i, to_place);

Finally the updated accumulator value is constructed, it is the last expression in the loop:

    (elem, new_count, new_result)

Following the loop body, as an argument to the loop, we initialize the accumulator in the following way.

  • set element prior to -1, in order to not match any other value.
  • set element count to 0.
  • set element result to 8 0's of size u3
}((u32:-1, u3:0, bits[8,3]:[u3:0, u3:0, u3:0, u3:0, u3:0, u3:0, u3:0, u3:0]));

And, finally, the function simply returns result:

  result
}

Testing

To test the two cases we've described above, we add the following two test cases right to this implementation file:

#![test]
fn test_prefix_scan_eq_all_zero() {
  let input = bits[8,32]:[0, 0, 0, 0, 0, 0, 0, 0];
  let result = prefix_scan_eq(input);
  assert_eq(result, bits[8,3]:[0, 1, 2, 3, 4, 5, 6, 7])
}

#![test]
fn test_prefix_scan_eq_doubles() {
  let input = bits[8,32]:[0, 0, 1, 1, 2, 2, 3, 3];
  let result = prefix_scan_eq(input);
  assert_eq(result, bits[8,3]:[0, 1, 0, 1, 0, 1, 0, 1])
}