Polyhedral Compilation Foundations

Lab 1: Computing the data set

• Review the slides here from the lecture #1.
• Review the slides here from the lecture #2.
• Review the slides here from the lecture #3.
• Review the slides here from the tutorial on ISCC (by Sven Verdoolaege).

Assignment for next week

Note: students will present their solution for the exercices on the computer during the class.

Exercise 1

You must first install iscc, that you can find in the Barvinok package.

For the following code, write the ISCC program in order to:

1. Represent the iteration domain of the statement.
2. Represent the data space of the program, for each array.
3. Count the number of points in the iteration domain.
4. Count the number of points in the data space.
5. Generate code to scan the iteration domain.
6. Generate code to scan the data space of the array A.

for (i = 0; i < N; ++i)
  for (j = 0; j < N; ++j)
    for (k = 0; k < N; ++k)
      C[i][j] += A[i][k] * B[k][j];

Lab 2: Tiling and other transformations with ISCC

• Review the slides here from the lecture #3.
• Review the slides here from the previous lecture.
• Review the slides here from the tutorial on ISCC (by Sven Verdoolaege).

Assignment for next week

Note: students will present their solution for the exercices on the computer during the class.

Exercise 1

Write an algorithm that creates automatically the (sequence of) transformations, in the form of a scattering function, to tile a loop nest of a given depth. The algorithm takes as an input (1) the number of loops d in a perfectly nested, permutable loop nest; (2) a tile size (scalar) s; and (3) a scattering function that maps some input space into a 2d+1 output space.

Exercise 2

Apply the algorithm of the previous exercise to compute the tiling transformation for matrix-multiply, where the input scattering permutes the loops i and k.

for (i = 0; i < N; ++i)
  for (j = 0; j < N; ++j)
    for (k = 0; k < N; ++k)
      C[i][j] += A[i][k] * B[k][j];

Lab 3: Communication generation

• Review the slides here from the previous lecture.
• Review the slides here from the tutorial on ISCC (by Sven Verdoolaege).
• Read the paper from Baskaran et al. Automatic data movement and computation mapping for multi-level parallel architectures with explicitly managed memories
• Read the paper from Alias et al. Program Analysis and Source-Level Communication Optimizations for High-Level Synthesis

Assignment for next week

Note: students will present their solution for the exercises on the computer during the class.

Exercise 1

Write the algorithm(s) to automatically create and place communications from main memory to a local buffer, for a program that is perfectly nested and contains only permutable loops. The program has to be tiled by your algorithm, and the communications must be at the granularity of a tile (that is, all elements required by a tile are copied before a tile executes).

Your algorithm must:

• Tile the original perfectly-nested loop nest (having d loops), using square tiles of size Ts
• Insert copy-in code for all read references
• Insert copy-out code for all written references
• Update the tile body to correctly index the local buffers
• Compute the buffer sizes

Exercise 2

Apply the algorithm of the previous exercise to tile and compute communications for matrix-multiply.

for (i = 0; i < N; ++i)
  for (j = 0; j < N; ++j)
    for (k = 0; k < N; ++k)
      C[i][j] += A[i][k] * B[k][j];

Exercise 3

Apply the algorithm of the previous exercise to tile and compute communications for jacobi-1d. A pre-transformation is required to make tiling legal: { s1[t,i] -> [t,2t+i,0]; s2[t,i] -> [t, 2t+i+1, 1]}.

for (t = 0; t < tsteps; t++)
  {
    for (i = 1; i < n - 1; i++)
	B[i] = 0.33333 * (A[i-1] + A[i] + A[i + 1]);
    for (i = 1; i < n - 1; i++)
	A[i] = B[i];
  }