Software Pipelining of Loops - 2024.1 English

AI Engine-ML Kernel and Graph Programming Guide (UG1603)

Document ID
UG1603
Release Date
2024-06-06
Version
2024.1 English

This section dives into software pipelining of loops. This is an important concept that enables the AI Engine-ML to concurrently execute different parts of a program. For example, a loop that requires a total of nine cycles to execute through one iteration is shown in the following figure, where sequential execution all the way to a full overlap pipelining is illustrated.

Figure 1. Pipelining Example

Counting the cycles through each of these examples, it is clear that the sequential execution requires 27 cycles to fully execute the three loop iterations, while the partially overlapped pipeline requires 13 cycles, and the fully pipelined loop requires only 11 cycles. From a performance perspective, it is therefore desirable to have a fully overlapping pipeline. However, this is not always possible, because resource constraints, as well as inter-iteration loop dependencies can prevent a full overlap (see the following figure).

Figure 2. Dependencies in Pipelining

In this example, the program performs load A (2 x 256-bit) in cycle 2, load B (2 x 256-bit) in cycle 3, and in cycle 6 and 7 it is executes operations on loop variable A. The remaining instructions of this iteration are of no importance with respect to the loop performance analysis.

Cycles 2 and 3 of this loop iteration execute 4 x 256-bit load operations. The required four loads are executed in two cycles because the AI Engine-MLs can only execute two loads per cycle. This is called a resource constraint. If the loop containing this iteration is supposed to be pipelined, this constraint limits the overlap to no less than two cycles. Similarly, code dependencies between iterations shown in cycle 6 and 7 can prevent additional overlap. In this case, the next iteration of the loop requires the value of A to be updated before it can be used by the loop, thus, limiting the overlap.

The AI Engine compiler reports on each loop in the following form.

Note: The core compilation report can be found in Work/aie/core_ID/core_ID.log and the -v option is needed to generate the verbose report.
HW do-loop #397 in "testbench.cc", line 132: (loop #16) :
Critical cycle of length 2 : b67 -> b68 -> b67
Minimum length due to resources: 2
Scheduling HW do-loop #397
(algo 1a)       -> # cycles: 9
(modulo)        -> # cycles: 2 ok (required budget ratio: 1)
(resume algo)     -> after folding:  2 (folded over 4 iterations)
  -> HW do-loop #397 in "testbench.cc", line 132: (loop #16) : 2 cycles
NOTICE: loop #397 contains folded negative edges
NOTICE: postamble created
Removing chess_separator blocks (all)

In the AI Engine compiler report shown previously, the section Critical cycle of length provides feedback on code dependencies, while the Minimum length due to resources indicates minimum overlap requirement due to resource constraints. The algo 1a line states the total amount of cycles for a single iteration. Given these numbers, there are a maximum of five iterations active at a time creating the pipeline.

The AI Engine compiler reports these five overlapping iterations (the current iteration plus four folded iterations) in the resume algo line. In addition, it states the initiation interval (II), the number of cycles a single iteration has to execute before the following iteration is started, which is two in this example.

In general, it is sufficient to provide the directive chess_prepare_for_pipelining to instruct the compiler to attempt software pipelining. When the number of loop iterations is a compile time constant, the chess compiler creates the optimum software pipeline.

In the case of a dynamic loop range (defined by a variable start/end), the compiler requires additional information to create an effective pipeline loop structure. This is performed through the directive chess_loop_range(<minimum>, <maximum>).

Note: If the number of cycles in the loop exceeds 64 cycles, pipelining can be disabled by the compiler for that loop. In this case, the compiler reports the following message.
(algo 1a)       -> # cycles: 167 (exceeds -k 64) -> no folding: 167
  -> HW do-loop #511 in "xxxx", line 794: (loop #8): 167 cycles
To increase the threshold for loop pipelining, the following option can be added to AI Engine compiler:
--Xchess="main:backend.mist2.maxfoldk=200"
Important: To generate the detailed report, it is important to enable verbose mode in Vitis IDE as shown in the following figure or use the -v option at the command line, as in, v++ -c --mode aie --aie -v.
Figure 3. Enabling Verbose Mode in Debug Reports

A modulo scheduling report can be generated for modulo scheduled loops by specifying the option -Xchess=main:backend.mist2.xargs=-ggraph for the AI Engine compiler. The modulo scheduling report is available for software pipelined loop with the name *_modulo.rpt in Work/aie/core_ID/Release/chesswork/<mangled_function_name>/*.rpt, where * is the block name. The modulo scheduling report also contains the information about register live ranges for register files, which may be useful to find inefficiencies in register assignment and can be improved by using chess_storage.

After compilation and linking is completed, you can open the compile log for an individual kernel in the Vitis IDE. For more information, see the AI Engine Tools and Flows User Guide (UG1076).