Array Accesses and Performance - 2021.1 English

The following code example shows a case in which accesses to an array can limit performance in the final RTL design. In this example, there are three accesses to the array mem[N] to create a summed result.

#include "array_mem_bottleneck.h"
 
dout_t array_mem_bottleneck(din_t mem[N]) {  

 dout_t sum=0;
 int i;

 SUM_LOOP:for(i=2;i<N;++i)
   sum += mem[i] + mem[i-1] + mem[i-2];
    
 return sum;
}

During synthesis, the array is implemented as a RAM. If the RAM is specified as a single-port RAM it is impossible to pipeline loop SUM_LOOP to process a new loop iteration every clock cycle.

Trying to pipeline SUM_LOOP with an initiation interval of 1 results in the following message (after failing to achieve a throughput of 1, Vitis HLS relaxes the constraint):

INFO: [SCHED 61] Pipelining loop 'SUM_LOOP'.
WARNING: [SCHED 69] Unable to schedule 'load' operation ('mem_load_2', 
bottleneck.c:62) on array 'mem' due to limited memory ports.
INFO: [SCHED 61] Pipelining result: Target II: 1, Final II: 2, Depth: 3.

The issue here is that the single-port RAM has only a single data port: only one read (or one write) can be performed in each clock cycle.

• SUM_LOOP Cycle1: read mem[i];
• SUM_LOOP Cycle2: read mem[i-1], sum values;
• SUM_LOOP Cycle3: read mem[i-2], sum values;

A dual-port RAM could be used, but this allows only two accesses per clock cycle. Three reads are required to calculate the value of sum, and so three accesses per clock cycle are required to pipeline the loop with an new iteration every clock cycle.

The code in the example above can be rewritten as shown in the following code example to allow the code to be pipelined with a throughput of 1. In the following code example, by performing pre-reads and manually pipelining the data accesses, there is only one array read specified in each iteration of the loop. This ensures that only a single-port RAM is required to achieve the performance.

#include "array_mem_perform.h"
 
dout_t array_mem_perform(din_t mem[N]) {  

 din_t tmp0, tmp1, tmp2;
 dout_t sum=0;
 int i;

 tmp0 = mem[0];
 tmp1 = mem[1];
 SUM_LOOP:for (i = 2; i < N; i++) { 
 tmp2 = mem[i];
 sum += tmp2 + tmp1 + tmp0;
 tmp0 = tmp1;
 tmp1 = tmp2;
 } 
    
 return sum;
}

Vitis HLS includes optimization directives for changing how arrays are implemented and accessed. It is typically the case that directives can be used, and changes to the code are not required. Arrays can be partitioned into blocks or into their individual elements. In some cases, Vitis HLS partitions arrays into individual elements. This is controllable using the configuration settings for auto-partitioning.

When an array is partitioned into multiple blocks, the single array is implemented as multiple RTL RAM blocks. When partitioned into elements, each element is implemented as a register in the RTL. In both cases, partitioning allows more elements to be accessed in parallel and can help with performance; the design trade-off is between performance and the number of RAMs or registers required to achieve it.