Loop tiling
Loop tiling (also known as loop blocking) is a program optimization technique whose aim is to improve locality of reference. It modifies the memory access pattern of the loop in order to reuse data already present in the data cache before that data gets evicted and needs to be reloaded from the memory.
The basic idea behind loop tiling is to split the processing of data into smaller segments called tiles or blocks. Because blocks are smaller than the original workset, they fit into data caches more readily.
Let's consider the following example:
for (int i = 0; i < n; i++) {
for (int j = 0; j < m; j++) {
c[i][j] = a[i] * b[j];
}
}
If you look more closely, the whole array b needs to be fetched from the
memory n times. This is not a problem if the size of the array b is small,
since then the values for b would be brought up from the main memory to the
data cache once and reused every time after that.
The problem happens when the array b is large. In that case, the CPU would
need to bring the values of the array b n times from the main memory,
since b doesn't fit the data cache.
The solution to this problem is loop tiling: instead of running the inner loop
over j from 0 to m and then increasing the variable i, we pick a constant
called TILE_SIZE and run the loop according to this pattern:
i | j |
|---|---|
i = 0 | j = 0, 1, ..., TILE_SIZE - 1 |
i = 1 | j = 0, 1, ..., TILE_SIZE - 1 |
... | ... |
i = n - 1 | j = 0, 1, ..., TILE_SIZE - 1 |
i = 0 | j = TILE_SIZE, TILE_SIZE + 1, ..., 2 * TILE_SIZE - 1 |
i = 1 | j = TILE_SIZE, TILE_SIZE + 1, ..., 2 * TILE_SIZE - 1 |
... | ... |
By running the loop like this, we are reusing the part of the array b which is
already in the data cache.
We can rewrite the code to use loop tiling:
for (int jj = 0; jj < m; jj += TILE_SIZE) {
for (int i = 0; i < n; i++) {
for (int j = jj; j < MIN(jj + TILE_SIZE, m); j++) {
c[i][j] = a[i] * b[j];
}
}
}
The careful reader might notice that after this intervention, the values for the
array a will be read m / TILE_SIZE times from the memory. If the size of
array a is large, then it can be useful to perform loop tiling on the loop
over i a as well, like this:
for (int ii = 0; ii < n; ii += TILE_SIZE_I) {
for (int jj = 0; jj < m; jj += TILE_SIZE_J) {
for (int i = ii; i < MIN(n, ii + TILE_SIZE_I); i++) {
for (int j = jj; j < MIN(m, jj + TILE_SIZE_J); j++) {
c[i][j] = a[i] * b[j];
}
}
}
}
Originally, data was processed inside a one large tile with dimensions n * m.
After loop tiling, data is processed inside many smaller tiles with dimensions
TILE_SIZE_I * TILE_SIZE_J. This kind of processing results in a better usage
of the memory subsystem.
Picking the values for tile size is typically done experimentally. Start off with some value, e.g. 16, and then change the value until the best performance is achieved.
When should loop tiling be applied?
There are two distinct scenarios where loop tiling is applied:
-
Iterating over the same dataset several times: this happens in the example above and we showed how loop tiling can help make the code run faster by shrinking the size of the workload in order to use the data while it is still in the data cache.
-
Cases with strided and random memory access patterns that cannot be improved with loop interchange because, with loop interchange, we fix one bad memory access pattern but introduce a new one: if the algorithm requires a strided memory access pattern for example, it is possible to optimize the loop with such a memory access pattern using loop tiling, even if there is no data reloading.
We already covered an example for (1) earlier, and loop tiling for (2) requires additional explanation.
Let's take as an example matrix transposition:
for (int i = 0; i < n; i++) {
for (int j = 0; j < n; j++) {
b[i][j] = a[j][i];
}
}
In this example, there is no data reusage. Each element of matrix a and matrix
b is accessed only once.
Under the assumption that the matrices are stored in row-major order, and
because of the principles of data locality, after accessing a[j][i] access to
a[j][i + 1], a[j][i - 1] and maybe a few other neighbors is very cheap
(since that data is in the data cache). But, access to a[j][i + 1] happens
much later after access to a[j][i] and doesn't benefit data locality.
Loop tiling can be used to decrease the time between consecutive accesses to
a[j][i] and a[j][i + 1]. We tile the loop like this:
for (int ii = 0; ii < n; ii += TILE_SIZE) {
for (int jj = 0; jj < n; jj += TILE_SIZE) {
for (int i = ii; i < MIN(n, ii + TILE_SIZE); i++) {
for (int j = jj; j < MIN(n, jj + TILE_SIZE); j++) {
b[i][j] = a[j][i];
}
}
}
}
Instead of accessing data inside one large tile with dimensions n * n, the
program is accessing data inside many smaller tiles with dimensions
TILE_SIZE * TILE_SIZE. This improves the loop's data locality and makes for a
faster loop.
What are the prerequisites for loop tiling?
To perform loop tiling, the loops have to be perfectly nested and a certain type of loops with loop-carried dependencies cannot be tiled.
Additionally, doing loop tiling only makes sense if there is data reuse or an inefficient memory access pattern that cannot be fixed with loop interchange. Without it, there can be no speed improvements.
How to perform loop tiling?
Loop tiling is typically performed on a loop nest. As an explanation it sounds quite complicated, but in practice there is a pattern that works most of the time. Let's take the example:
for (int i = 0; i < n; i++) {
for (int j = 0; j < n; j++) {
b[i][j] = a[j][i];
}
}
There are two loops in this loop nest. For each loop we introduce another outer
loop. i.e. for loop over i we introduce another loop over ii and for loop
over j we introduce another loop over jj.
The outer loops (over ii and jj) iterate from 0 to n with an increment of
TILE_SIZE. The corresponding inner loop (over i and j) iterates from the
value of the corresponding outer loop index (ii for i and jj for j) to
the length of the tile or end of the array. It looks like this:
| Outer loop | Inner loop |
|---|---|
for (int ii = 0; ii < n; ii += TILE_SIZE) | for (int i = ii; i < MIN(ii + TILE_SIZE, n); i++) |
for (int jj = 0; jj < n; jj += TILE_SIZE) | for (int j = jj; j < MIN(jj + TILE_SIZE, n); j++) |
We first place outer loops and then inner loops. The resulting code looks like this:
for (int ii = 0; ii < n; ii += TILE_SIZE) {
for (int jj = 0; jj < n; jj+=TILE_SIZE) {
for (int i = ii; i < MIN(n, ii +TILE_SIZE); i++) {
for (int j = jj; j < MIN(n, jj + TILE_SIZE); j++) {
b[i][j] = a[j][i];
}
}
}
}