Array Block and Cyclic Partitioning

This example shows how to use array block and cyclic partitioning to improve performance of a kernel

KEY CONCEPTS: Kernel Optimization, Array Partitioning, Block Partition, Cyclic Partition

KEYWORDS: #pragma HLS ARRAY_PARTITION, cyclic, block, factor, dim

This example demonstrates how to use array block and cyclic partitioning to improve the performance of the kernel. Matrix multiplication is performed in this example which would require repeated access to rows of former matrix and columns of latter. Array can be partitioned across different dimensions to reduce the latency of these transfers.

#pragma HLS ARRAY_PARTITION is used to partition an array into multiple smaller arrays with more number of ports for read and write operations. This results in improved throughput of the design.

Arrays can be partitioned in three ways, cyclic where elements are put into smaller arrays one by one in the interleaved manner untill the whole array is partitioned, block where elements are put into smaller arrays from continuous blocks of original array(number of smaller arrays is defined by factor) and complete where array is decomposed into individual elements each having own read/write ports.

#pragma HLS ARRAY_PARTITION variable = A dim = 1 cyclic factor = 16
#pragma HLS ARRAY_PARTITION variable = B dim = 1 block factor = 16

Following is the log reported while running the design on U250 platform:

Found Platform
Platform Name: Xilinx
INFO: Reading ./build_dir.hw.xilinx_u250_gen3x16_xdma_4_1_202210_1/matmul.xclbin
Loading: './build_dir.hw.xilinx_u250_gen3x16_xdma_4_1_202210_1/matmul.xclbin'
Trying to program device[0]: xilinx_u250_gen3x16_xdma_shell_3_1
Device[0]: program successful!
|-------------------------+-------------------------|
| Kernel                  |    Wall-Clock Time (ns) |
|-------------------------+-------------------------|
| matmul: naive           |                   20496 |
| matmul: partition       |                     866 |
|-------------------------+-------------------------|
| Speedup                 |                  23.664 |
|-------------------------+-------------------------|
Note: Wall Clock Time is meaningful for real hardware execution only, not for emulation.
Please refer to profile summary for kernel execution time for hardware emulation.
TEST PASSED

EXCLUDED PLATFORMS:

• All NoDMA Platforms, i.e u50 nodma etc

• Versal DFX platform

DESIGN FILES

Application code is located in the src directory. Accelerator binary files will be compiled to the xclbin directory. The xclbin directory is required by the Makefile and its contents will be filled during compilation. A listing of all the files in this example is shown below

src/host.cpp
src/matmul.cpp
src/matmul_partition.cpp

Access these files in the github repo by clicking here.

COMMAND LINE ARGUMENTS

Once the environment has been configured, the application can be executed by

./partition_cyclicblock <matmul XCLBIN>