Loop Unrolling

What is loop unrolling?

Now that we have discussed loop pipelining, we can get even more saving with unrolling.
Unrolling duplicates the loop body to perform multiple iterations in parallel. For example, an unroll factor of 4 is equivalent to a loop like:

for (i=0; i < n; i += 4) {
    c_buf[i] = a_buf[i] * b_buf[i];
    c_buf[i+1] = a_buf[i+1] * b_buf[i+1];
    c_buf[i+2] = a_buf[i+2] * b_buf[i+2];
    c_buf[i+3] = a_buf[i+3] * b_buf[i+3];
}

where the four multiplies are performed simultaneously in each iteration. This unrolling is realized in hardware by instantiating up to four physical multiply units that run in parallel. Ideally, this parallelism will reduce the latency by a factor of 4.

In the above example, suppose n=1024 and we unroll by a factor of 4. We say n=1024 is the number of iterations and the loop has 256 trips — each trip performs 4 multiplications in parallel due to unrolling.

In Vitis HLS, you do not need to manually unroll the loop. You can simply add the pragma:

#pragma HLS UNROLL factor=4

Array partitioning

Unforunately, unrolling may not get the full gain. In addition to multiplications, each iteration must read from buffers a_buf and b_buf and store to c_buf. Generally each of these buffers can only support one or two accesses per cycle. Hence, they will be a bottleneck. To avoid this bottleneck, we partition the buffers into discrete blocks, each with its own addressing logic to parallelize access. This is done with the compiler directive:

#pragma HLS ARRAY_PARTITION variable=a_buf type=cyclic factor=4  dim=1

Note:

• The parameter dim=1 refers to partitioning along the first dimension — the index used in a_buf[i].
• We used type=cyclic to partion so that it puts 0,4,8,... in the first partition, 1,5,9,... in the second partition and so on.

Synthesis with unrolling

To enable unrolling, say by a factor of 4: In include/vmult.h, set:

#define UNROLL_FACTOR 4

Then re-run synthesis and examine the loop report. You should see:

• Interval = 1 meaning II=1
• Trip count = 256 meaning that loop with n=1024 iterations only required 256 trips.

Thus, we reduce the timing by an additional factor of 4.

---

Go to Automated parameter sweeping.