CPRI Compression

CPRI Compression

CPRI bandwidths needs have increased rapidly over the last few years; this white paper shows a compression algorithm with by 2:1 compression and minimal impact on signal fidelity using an insignificant amount of logic.

Typical LTE macro radio basestation systems are composed of two parts, the baseband processing and the radio. The interface between these two functional blocks is defined by a publically available specification called CPRI (Common Packet Radio Interface). In this specification, these blocks are defined as the REC (Radio Equipment Controller) and the RE (Radio Equipment). CPRI defines various topologies including point to point, point to multi-point, chains and rings. CPRI transports synchronization, C&M (control and management) and baseband IQ data. Introduction

8/8b/RECRE.JPG

CPRI was originally developed for UMTS, however, as wireless standards have evolved, the bandwidth needs for IQ data has dramatically increased.

1/1f/CPRIBW.JPG

The increase in sample rates, number of antennas and the number of radios has driven the CPRI standard to almost double in bandwidth every 2 years. This drives up the cost of digital implementation (logic and transceivers) and the optics (laser modules and fibers). The recent CPRI V6.0 version of the standard has introduced 10.1376Gbps links and also increased the efficiency of the link for this rate by using 66b64b encoding instead of the 8b10b used on other line rates. Additional efficiency is highly desirable.

8/88/CPRIoverTime.JPG

This paper describes a means of compressing the IQ data in order to halve the transferred data for a relatively small loss in fidelity.

Mu-Law & A-Law Compression

1/18/MUEqn.JPG

Mu-Law & A-Law compression is often used in audio compression schemes. It works by expanding the dynamic range of the signal so that quantization loss effects are minimized. Mu-Law compression is specified in (ITU-T Recommendation G.711 and G.191).

The rate of expansion is controlled through the selection of a constant Mu_compand_val. In these audio systems, a high value of Mu_compand_val (255) is specified. It yields a 2^n exponent and piecewise linear approximation through direct bit shifting where the amount of shifting is determined by the exponent (see table below). Such a high value of Mu_compand_val maps an exponentially increasing number of samples to the same exponent and yields very poor results for CPRI compression – a shallower exponent/ rate of expansion is required (see graph below).

A much lower value of Mu_compand_val is used in this example. Note that a Mu_compand_val of 1 is equivalent to the round(x) function.

b/b7/MuApprox.png

Without any approximation or quantization effects, when Mu_compand_val reaches 128 the EVM has grown to >1%.

FPGA Implementation

Choosing a Mu of 255 is an efficient implementation point but poor for CPRI waveforms. Natural logarithmic operations are resource intensive functions on FPGAs. The results shown below are generated from an FPGA implementation where segmented offset a linear approximations are used (patent pending). The compressed value is represented by an exponent and mantissa. Whilst the number of bits In this example, a logarithmic approximation is constructed using 3 segments where each segment is identified using 2 (exponent) bits. The first segment covers the first quarter of the waveform, the second segment the second quarter and the last segment, the second half. These segment transitions are on binary boundaries. The plot below shows the idealized Mu-law and the corresponding linear approximation.

1/1d/Histogram.png

CPRI waveforms are close to Gaussian in distribution; this means that the majority of samples are located around the zero crossing. Mu-Law compression pushes the distribution away from zero and forms a much more even distribution. The implemented version uses a segmented approximation; this is characterized in the resulting distribution as ‘kinks’.

Test Scenario

Signal fidelity in LTE cellular radio systems is defined by 3GPP in TS 36.104 by EVM (Error Vector Measurement). The measurement point is following removal of the CP (Cyclic Prefix), performing a FFT and performing per-carrier amplitude/phase correction.

EVM is the size of the vector from the ideal constellation point to the measured one. For a 64QAM modulated signal, the LTE specification is 8%.

Using Agilent test equipment, a LTE test model TM3.1 was used with different compression techniques applied. The waveforms were played from a MXG into an MXA running VSA LTE demodulation personalities to measure EVM as per the 3GPP standard.

8/83/MuMXA.png

The following plot shows the relative EVM degradation comparing no compression (equipment noise floor), simple quantization and the implemented Mu-Law approximation. A 16-bit input is used and then compressed to a differing number of bits. Mu-Law compression provides approximately 1-bit of additional fidelity compared to rounding.

9/9f/ByNumBits.JPG

Implementation

The following table shows the resource utilization in an Altera FPGA. Latency is 4 clock cycles.

	ALUTs	Registers	Multipliers	Memories
Mu-Law Compressor	12	46	0	0
Mu-Law Decompressor	22	26	0	0

Conclusion

Using Altera's segmented linear approximation of Mu-Law compression, CPRI users can achieve 2:1 compression from 16 to 8 bits using a tiny amount of logic. The loss in signal fidelity is less than 0.8% EVM. For more information contact ssewireless@altera.com.