FFT Dynamic Point Stream - 2022.2 English - UG1483

Vitis Model Composer User Guide (UG1483)
Document ID
UG1483
Release Date
2023-01-13
Version
2022.2 English


Library

AI Engine/DSP/Stream IO

Description

FFT Dynamic Point Stream implementation targeted for AI Engines. This block supports performing different lengths of FFT on input data streams. Each frame must be preceded by a 256-bit header which describes run-time parameters of that frame. The format of the header is described below. The block has two stream ports per subframe processor to maximize performance. The specified value for SSR parameter should be of the form 2^N, where N is a positive integer.

Note:
  • Every input frame must be appended by a header.
  • If the FFT size (N) is smaller than the frame size, the block takes the N point FFT of the first N samples in the input frame, and discards the rest of the samples in the frame.
  • If the frame size is an integer multiple (M) of max FFT size, M FFT iterations will be performed on a given input frame, resulting in multiple iterations of output samples, reducing the number of times the kernel needs to be triggered to process a given number of input data samples. As a result, the overheads incurred during kernel triggering are reduced and overall performance is increased.
Figure 1. Header Format
Table 1. Header Description
Index Field Name Description
1 Direction
  • 0 = Inverse FFT
  • 1= Forward FFT
2 Point size (radix 2 stages) Point size described as power of 2. For example, value '5' describes a point size of '32'
3 Reserved Reserved
4 Status (output only)
  • 0 = Legal point size
  • 1 = Illegal point size

Example Header Format

To implement the 32-point size FFT of cint16 datatype, the format of the header that should be preceded to the input data window should be as shown in the following figure.

Figure 2. Example Header Format

Example (Input to Dynamic Point FFT)

To implement the FFT of different sizes, say, 64, 32 and 128, of type cint32, the format of the input should be as follows:

complex([1 6 0 0 ones(1,64) 1 5 0 0 ones(1,32) 1 7 0 0 ones(1,128)])

Parameters

Main
Input/Output Data Type:
  • Describes the type of individual data samples input/output of the dynamic point FFT. It can be cint16, cint32, cfloat types.
FFT Maximum Size:
  • Specifies the maximum FFT size that is supported by Dynamic Point FFT. You can perform different lengths of FFT on different input data frames. It must be a power of 2 with a minimum value of 16. The maximum value supported by the library element is 65536.
Input Window Size:
  • Specifies the number of samples in the input frame excluding the header. The value must be in the range 16 to 65536 and the default value is 60.
Scale Output down by 2^:
  • Describes the power of 2 shift down applied before output.
SSR:
  • This parameter is intended to improve performance and support FFT sizes beyond the limitations of a single tile. For an SSR value of 'n' (which must be of the form 2^N, where N is a positive integer), the FFT operation is performed in parallel and the actual FFT size is divided by 'n'. For example, a 16384 point FFT with an SSR value of 8 creates 8 stream inputs and there will be 8 subframe FFTs each of point size 2048. The specified FFT size and SSR values should be such that FFT size / SSR should not exceed 2048.
Advanced
Target Output Throughput (MSPS):
  • Specifies the output sampling rate of the FFT function in Mega Samples per Second (MSPS). The value must be in the range 1 to 1000 and the default value is 200.
Specify the Number of Cascade Stages:
  • When this option is disabled, the tool will determine the FFT configuration that best achieves the specified input sampling rate. When the option is enabled, the Number of cascade stages can be specified (which describes the number of AI Engine processors to split the operation over). This allows resources to be traded for higher performance, but the specified input sampling rate constraint may not be achieved. The value must be in the range of 1 to 9.