Application Report

SPRA947A June 2009

Signal Processing Examples Using the TMS320C67x

Digital Signal Processing Library (DSPLIB)
Anuj Dharia & Rosham Gummattira

TMS320C6000 Software Applications

ABSTRACT

The TMS320C67x digital signal processing library (DSPLIB) provides a set of C-callable,
assembly-optimized functions commonly used in signal processing applications, e.g.,
filtering and transform. The DSPLIB includes several functions for each processing category,
based on the input parameter conditions, to provide parameter-specific optimal performance.
Therefore, it is important to understand the differences and requirements of the functions in
each category. This application report presents the usage and performance of three key
signal processing categories, i.e., finite impulse response (FIR), bi-quadratic infinite impulse
response (IIR), and fast Fourier transform (FFT), to help users better utilize DSPLIB in their
system development.

Contents
1

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

Benchmarking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1 Emulation/Simulation Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2 Cycle Count Measurement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3 Example Scenario and Expected Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
3.1 Finite Impulse Response (FIR) Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
3.1.1 DSPF_sp_fir_gen Single Precision Generic FIR filter . . . . . . . . . . . . . . . . . . . . . . . . . . 7
3.1.2 DSPF_sp_fir_r2 Single Precision Radix 2 FIR filter . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
3.1.3 DSPF_sp_fir_cplx Single Precision Complex Radix 2 FIR filter . . . . . . . . . . . . . . . . . 7
3.1.4 FIR Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3.2 Infinite Impulse Response (IIR) Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.2.1 DSPF_sp_biquad Single Precision Bi-quadratic IIR filter . . . . . . . . . . . . . . . . . . . . . . 10
3.2.2 IIR Filter Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.3 Fast Fourier Transform (FFT) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.3.1 DSPF_sp_cfftr2_dit Complex Radix 2 FFT using Decimation-In-Time . . . . . . . . . . 13
3.3.2 DSPF_sp_cfftr4_dif Complex Radix 4 FFT using Decimation-In-Frequency . . . . . 13
3.3.3 DSPF_sp_fftSPxSP Cache Optimized Mixed Radix FFT with digit reversal . . . . . . 13
3.3.4 FFT Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.3.5 Performance Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

3
3
4
5

Trademarks are the property of their respective owners.

1

SPRA947A

List of Figures
Figure 1
Figure 2
Figure 3
Figure 4
Figure 5
Figure 6
Figure 7
Figure 8
Figure 9
Figure 10

Memory Hierarchy and Potential Overhead . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

C6713 Linker Command File . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
Frequency Response of a Low-pass FIR Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
FIR Filter Input . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
FIR Filter Output . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
Frequency Response of Low-Pass IIR Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
IIR Filter Input . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
IIR Filter Output . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
512-Point FFT Input . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
Magnitude of the Output . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
List of Tables

Table 1
Table 2
Table 3
Table 4
Table 5
Table 6
Table 7
Table 8

C6713 DSK Key Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

Stall Cycles related to L1D . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
FIR Filter Design Specifications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
FIR Filter Benchmarks (240 Kernel Coefficients (nh) and 200 Output Samples (nr)) . . . . 9
IIR Filter Design Specifications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
IIR Filter Benchmarks (200 output samples (nx)) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
Single-Pass FFT Benchmarks for 1024 Point input . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
Single-Pass vs. Multi-Pass FFT Benchmarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

Introduction
TMS320C67x is an advanced very long instruction word (VLIW) processor well suited for
real-time signal processing applications with its high computing power and large on-chip
memory. It also provides enhanced direct memory access (EDMA) and cache to efficiently
transfer data to/from off-chip memory/device. To help users shorten the time-to-market in system
development, we provide a set of assembly-optimized functions named digital signal processing
library (DSPLIB). Each function in the DSPLIB is designed to provide the best performance
possible by optimally utilizing available resources and avoiding potential resource conflicts.
The DSPLIB includes several functions for each processing category, based on the input
parameter conditions to provide parameter-specific optimal performance. When users utilize
DSPLIB, therefore, it is important to understand the differences and requirements of the
functions in each category.
It is also important to understand potential overhead related to memory hierarchy to estimate
and improve the actual performance of a system being developed, Figure 1 shows the memory
hierarchy of C67x and related potential overhead. For example, when new code needs to be
fetched and/or the whole program does not fit the level-one program cache (L1P), L1P cache
misses can occur, stalling the CPU until the required code is fetched. Similarly, when the whole
data do not fit the level-one data cache (L1D) and/or a new set of data needs to be transferred
to/from off-chip memory/device, L1D cache misses stall the CPU. All L1P and L1D misses are
serviced by the level-two cache/SRAM (L2 cache/SRAM).

Signal Processing Examples Using the TMS320C67x Digital Signal Processing Library (DSPLIB)

SPRA947A
C67x
CPU

L1P

L1 program cache misses

L1 data cache misses

L2 cache misses

Off-chip memory accesses

L1D

L2 Cache/SRAM

4
DMA/EMIF

Off-chip memory

Figure 1. Memory Hierarchy and Potential Overhead

Similar to the L1D misses, L2 cache misses occur if the whole code and data do not fit the L2
cache and/or a new set of data needs to be transferred to/from off-chip memory/device. The L2
miss overhead can be significant compared to the L1P/L1D miss overhead because the L2
cache needs to communicate with slow off-chip memory/device.
L2 SRAM can also be used to service L1D/L1P misses with DMA to transfer code/data between
L2 SRAM and off-chip memory/device. The data transfer with DMA is typically more effective
than that with L2 cache due to its nature of longer burst transactions, thereby reducing memory
access latency overhead. However, the DMA transfer can involve more programming effort
because data transfers and synchronization have to be manually managed. TMS320C67x
provides both cache and DMA mechanisms to allow users to choose a right mechanism
depending on situations.
This application note presents the usage and performance of three key categories, i.e., finite
impulse response (FIR) filter, infinite impulse response (IIR) filter and fast Fourier transform
(FFT), to help TMS320C67x users utilize DSPLIB in system development.

Benchmarking

2.1

Emulation/Simulation Setup
A TMS320C6713 DSP starter kit (DSK) is used in this application report to measure cycle
counts. Table 1 lists key features of the C6713 DSK, which are important factors in the
performance analysis and optimization. More details on the C6713 internal memory structure
and operations can be found in the TMS320C621x/C671x DSP Two-Level Internal Memory
Reference Guide (SPRU609).

Signal Processing Examples Using the TMS320C67x Digital Signal Processing Library (DSPLIB)

SPRA947A

Table 1. C6713 DSK Key Features

Item

Description

Clock frequency

225 MHz

L1P

4-kbyte, direct-mapped, 64-byte cache line

L1D

4-kbyte, 2-way set associative, 32-byte cache line, 64-bit wide dual-ported

L2 SRAM

5-cycle L1P miss penalty, 4-cycle L1D miss penalty, up to 64 Kbytes, four 64-bit
banks

L2 cache

5-cycle L1P miss penalty, 4-cycle L1D miss penalty, up to 64 Kbytes, 1/2/3/4-way
set associative, 128-byte cache line, four 64-bit banks

L2 to L1D read path

128 bits

L1D to L2 write buffer

32-bit, 4-entry, L2 can process a write request every 2 cycles

EMIF

32-bit bus

The C6713 DSK is connected to a PC using a USB A/B connector cable. If you use simulation,
select C67xx Cycle Accurate Simulator. The cycle counts obtained from simulation might not
be accurate because the simulator ignores L1/L2 cache misses and off-chip memory accesses.
Software version numbers used in this application report are as follows:

2.2

Code Composer Studio version 2.2

C67x DSPLIB version 1.0

Cycle Count Measurement

The built-in timer in C6713 is used to measure cycle counts for DSPLIB examples. The following
sample code shows how to set up the timer and measure the cycle counts with the Chip Support
Library (CSL).
hTimer = TIMER_open(TIMER_DEVANY,0);

/* open a timer */

/**/
/* Configure the timer. 1 count corresponds to 4 CPU cycles in C67 */
/**/
/* control

period

initial value

*/

TIMER_configArgs(hTimer, 0x000002C0, 0xFFFFFFFF, 0x00000000);

/* */
/*

Compute the overhead of calling the timer.

*/

/* */

start

= TIMER_getCount(hTimer); /* to remove L1P miss overhead */

start

= TIMER_getCount(hTimer); /* get count */

stop

= TIMER_getCount(hTimer); /* get count */

Signal Processing Examples Using the TMS320C67x Digital Signal Processing Library (DSPLIB)

SPRA947A

overhead = stop start;

start = TIMER_getCount(hTimer);

/* get count */

/* */
/*

Call a function here.

*/

/* */
diff = (TIMER_getCount(hTimer) start) overhead;

/* get count */

TIMER_close(hTimer);
printf(%d cycles \n, diff*4);

The maximum resolution of the timer is 4 CPU cycles since the input to the timer is fixed to the
CPU clock, divided by four. The function call overhead for the TIMER_getCount() is roughly
measured and compensated. Additional information on the timer registers can be found in
TMS320C6000 Peripherals Reference Guide (SPRU190).

2.3

Example Scenario and Expected Performance

To analyze the potential overhead related to the memory hierarchy, data is stored in the L2
SRAM of the chip. Figure 2 shows a linker command file used to store data on chip.
MEMORY
{
L2SRAM:

o = 00000000h

l = 00010000h

/* 64 kbytes */

}
SECTIONS
{
.cinit

>

L2SRAM

.text

>

L2SRAM

.stack

>

L2SRAM

.bss

>

L2SRAM

.const

>

L2SRAM

.data

>

L2SRAM

.far

>

L2SRAM

.switch

>

L2SRAM

.sysmem

>

L2SRAM

.tables

>

L2SRAM

.cio

>

L2SRAM

Figure 2. C6713 Linker Command File

The overhead with off-chip memory accesses is not presented in this report because the
overhead can be minimized by overlapping the data transfer time and the computations time
with the EDMA.

Signal Processing Examples Using the TMS320C67x Digital Signal Processing Library (DSPLIB)

SPRA947A

Additionally, a scenario in which data are already in L1D is not presented in this report because
these cycle counts are very close to the formula cycle counts listed in the TMS320C67x DSP
Library Programmers Reference (SPRU657).
When data in on chip in the L2 SRAM, L1D miss overhead needs to be considered. Table 2 lists
expected stall cycles related to L1D read and/or write transactions. When there are read
transactions only, the number of stall cycles is the number of L1D read misses times the L1D
miss penalty (i.e. 4 cycles). In case of write transactions only, there is no stall unless the write
buffer is full. The write buffer is 32-bit wide, and allows up to four outstanding misses.
When there are both read and write transactions, the L1D read miss penalty can increase
because, to maintain data coherency, the write buffer is flushed before a read miss is serviced.
Table 2. Stall Cycles related to L1D

Transaction

Number of stall cycles

Read transaction only

Number of L1D read misses * L1D miss penalty.

Write transaction only

No stall cycle unless the write buffer is full.

Read and write transactions

Number of L1D read misses * (L1D miss penalty + additional cycles for
write buffer flush)

Examples
This section presents the usage and performance of three key signal processing categories, i.e.,
finite impulse response (FIR) filter, infinite impulse response (IIR) filter and fast Fourier transform
(FFT). To minimize the variation in cycle count measurement, be sure to select the Reset menu
(under Debug in Code Composer Studio) before running an example.

3.1

Finite Impulse Response (FIR) Filter

A generalized FIR filter of N filter coefficients, h(k), is defined as:

h(k) x(n * k)

N*1

y(n) +

k+0

The C67x DSPLIB provides 3 FIR functions:

DSPF_sp_fir_gen
DSPF_sp_fir_r2
DSPF_sp_fir_cplx

The prototype and requirements of the FIR functions follow.

3.1.1

DSPF_sp_fir_gen Single Precision Generic FIR filter

void DSPF_sp_fir_gen(const float * restrict x, const float * restrict h, float
* restrict r, int nh, int nr)

x points to a floating point array of length nr+nh1 which holds the input samples.

r points to a floating point array of length nr which holds the outputs.

h points to a floating point array of length nh which holds the coefficients. The coefficients
need to be placed in h in reverse order.

Signal Processing Examples Using the TMS320C67x Digital Signal Processing Library (DSPLIB)

SPRA947A

nh is the number of coefficients. nh must be greater than or equal to 4.

nris the number of outputs. nr must be greater than or equal to 4.

3.1.2

DSPF_sp_fir_r2 Single Precision Radix 2 FIR filter

void DSPF_sp_fir_r2(const float *restrict x, const float *restrict h, float
*restrict r, int nh, int nr)

x points to a floating point array of length nr+nh1 which holds the input samples. x must be
double word aligned and padded with 4 extra words at the end.

h points to a floating point array of length nh which holds the coefficients. The coefficients
need to be placed in h in reverse order. h must be double word aligned and padded with 4
extra words at the end.

r points to a floating point array of length nr which holds the outputs.

nh is the number of coefficients. nh must be a multiple of 2 and greater than or equal to 8.

nr is the number of outputs. nr must be a multiple of 2 and greater than or equal to 2.

3.1.3

DSPF_sp_fir_cplx Single Precision Complex Radix 2 FIR filter

void DSPF_sp_fir_cplx(const float *restrict x, const float *restrict h, float
*restrict r, int nh, int nr)

3.1.4

xpoints to a floating point array of length 2*(nr+nh1) which holds the input samples. x must
be double word aligned and point to the 2*(nh1)th element (&x[2*nh1]).

h points to a floating point array of length 2*nh which holds the coefficients. The coefficients
need to be placed in h in normal order. h must be double word aligned.

r points to a floating point array of length 2*nr which holds the outputs.

nh is the number of coefficients. nh must be greater than or equal to 5.

nr is the number of outputs. nr must be a multiple of 2 and greater than or equal to 2.

FIR Example
This example demonstrates the use of the C67x DSPLIB FIR filtering capabilities. First, filter
coefficients are generated in Matlab using the Filter Design and Analysis Tool (Matlab command:
fdatool) with filter specifications listed in Table 3. The frequency response of this FIR filter is
shown in Figure 3.
Table 3. FIR Filter Design Specifications
Filter Type

Low-pass

Order

239 (240 for DSP_fir_sym)

Design Method

Window (Kaiser with a beta of 0.5)

Sampling frequency

44,100 Hz

Cut-off frequency

10,000 Hz

Signal Processing Examples Using the TMS320C67x Digital Signal Processing Library (DSPLIB)

SPRA947A

Figure 3. Frequency Response of a Low-pass FIR Filter

Input data to the FIR filter are generated in floating point format as follows:
x[i] = (sin(2 * PI * F1 * i / Fs) + sin(2 * PI * F2 * i / Fs));

Where F1 and F2 are the two input data frequencies and Fs is the sampling frequency (44,100
Hz). Figure 4 and Figure 5 show the results of the FIR filter. Figure 4 shows a sinusoidal input
described earlier where F1 = 370 Hz and F2 = 10500 Hz. Since 10,500 Hz is above the cut-off
frequency, this frequency is attenuated and only a 370 Hz sinusoidal wave remains as shown in
Figure 5.

Figure 4. FIR Filter Input

Signal Processing Examples Using the TMS320C67x Digital Signal Processing Library (DSPLIB)

SPRA947A

Figure 5. FIR Filter Output

Table 4 lists the performance for the three FIR functions. As expected, the performance is best
for the radix 2 FIR because its more stringent restrictions allow for better loop unrolling and
software pipelining. The complex FIR takes significantly longer than the rest.
Table 4. FIR Filter Benchmarks (240 Kernel Coefficients (nh) and 200 Output Samples (nr))
Number of Cycles
Functions

Formula

Observed (with data in L2 SRAM)

DSPF_sp_fir_gen

24508 = (4*floor((nh1)/2)+14)*(ceil(nr/4)) + 8

24968

DSPF_sp_fir_r2

24034 = (nh* nr)/2 + 34

24528

DSPF_sp_fir_cplx

96033 = 2* nh * nr + 33

96864

With the data in the L2 SRAM the observed cycle count is very similar to the formula cycle
count. The discrepancy is realized when call overhead, L1D read misses, and L1D write misses
are taken into account.

3.2

Infinite Impulse Response (IIR) Filter

The DSPLIB offers a 2nd order bi-quadratic IIR filter defined as:
y(n) +

b(k) x(n * k) * a(k)y(n * k)

2

k+0

k+1

where x(n) and y(n) are the input and output data, and a(k) and b(k) are the filter coefficients.
The a(k) are auto-regressive (AR) coefficients (poles of the transfer function). The b(k) are
moving-average (MA) coefficients (zeros of the transfer function).
IIR filters generally have nonlinear phase responses, but can meet magnitude response
specifications with much lower orders than FIR filters. However, due to their nature of instability,
care must be taken in their design to meet stability criteria.
The prototype and requirements of the IIR function follows.

Signal Processing Examples Using the TMS320C67x Digital Signal Processing Library (DSPLIB)

SPRA947A

3.2.1

DSPF_sp_biquad Single Precision Bi-quadratic IIR filter

void DSPF_sp_biquad (float* x, float* b, float* a, float* delay, float* r, int
nx)

3.2.2

x points to a floating point array of length nx which holds the input samples.

b points to a floating point array of length 3 which holds the MA coefficients: b[0], b[1], and
b[2]. b must be double-word aligned.

a points to a floating point array of length 2 which holds the AR coefficients: a[0] and a[1]. a
must be double-word aligned.

delay points to a floating point array of length 2 which holds the delay coefficients: d[0] and
d[1]

r points to a floating point array of length nx which holds the output samples.

nx is the length of the coefficient array. nx must be greater than or equal to 4.

IIR Filter Example

This example demonstrates the use of the C67x DSPLIB IIR filtering capabilities. First, filter
coefficients are generated in Matlab using the Filter Design and Analysis Tool (Matlab command:
fdatool) with filter specifications listed in Table 5. The frequency response of this FIR filter is
shown in Figure 6.
Table 5. IIR Filter Design Specifications
Filter Type

Low-pass

Order

Design Method

Butterworth

Sampling frequency

44,100 Hz

Cut-off frequency

8,000 Hz

Figure 6. Frequency Response of Low-Pass IIR Filter

10

Signal Processing Examples Using the TMS320C67x Digital Signal Processing Library (DSPLIB)

SPRA947A

Input data to the IIR filter are generated in floating point format as follows:
x[i] = (sin(2 * PI * F1 * i / Fs) + sin(2 * PI * F2 * i / Fs));

where F1 and F2 are the two input data frequencies and Fs is the sampling frequency
(44,100 Hz). Figure 7 and Figure 8 show the results of the FIR filter. Figure 7 shows a sinusoidal
input described earlier where F1 = 370 Hz and F2 = 18500 Hz. Since 18,500 Hz is above the
cut-off frequency, this frequency is attenuated and only a 370 Hz sinusoidal wave remains as
shown in Figure 8.

Figure 7. IIR Filter Input

Figure 8. IIR Filter Output

Table 6 shows the performance of the IIR filter.
Table 6. IIR Filter Benchmarks (200 output samples (nx))
Number of Cycles
Functions

Formula

DSPF_sp_biquad

876 = 4* nx + 76

Observed (with data in L2 SRAM)

1196

Signal Processing Examples Using the TMS320C67x Digital Signal Processing Library (DSPLIB)

11

SPRA947A

With the data in the L2 SRAM, the observed cycle count is very similar to the formula cycle
count. The discrepancy is realized when call overhead, L1D read misses, and L1D write misses
are taken into account.

3.3

Fast Fourier Transform (FFT)

FFT is widely used for frequency-domain processing and spectrum analysis. It is a
computationally efficient discrete Fourier transform (DFT) defined as:

xn WknN, k + 0,..., N * 1

N*1

X(k) +

n+0

where
*2j pnkN
W kn
N +e

The C67x DSPLIB provides three FFT functions:

1. DSPF_sp_cfftr2_dit
2. DSPF_sp_cfftr4_dif
3. DSPF_sp_fftSPxSP
The C67x DSPLIB also provides two inverse FFT functions:
1. DSPF_sp_icfftr2_dit (used for both radix 2 and radix 4 FFTs)
2. DSPF_sp_ifftSPxSP
The prototypes and restrictions for the FFT functions follow:

3.3.1

DSPF_sp_cfftr2_dit Complex Radix 2 FFT using Decimation-In-Time

void DSPF_sp_cfftr2_dit(float * x, float * w, short n)

x points to a floating-point array of length 2*n which holds n complex input samples. x must
be double-word aligned. After running the function, the output will also be stored in x. The
output must be bit-reversed using the bit reverse function found in the FFT support: bit_rev.

w points to a floating-point array of length n which holds the n/2 twiddle factors. w can be
created with the radix 2 twiddle generation function found in the FFT support: tw_genr2fft.
After creating the array, w must be bit-reversed using bit_rev.

n is the length of the FFT in complex samples. n must be a power of 2 and greater than or
equal to 32.

3.3.2

DSPF_sp_cfftr4_dif Complex Radix 4 FFT using Decimation-In-Frequency

void DSPF_sp_cfftr4_dif(float * x, float * w, short n)

12

x points to a floating point array of length 2*n which holds n complex input samples. After
running the function, the output will also be stored in x. The output must be digit-reversed
using the digit reverse functions found in the FFT support: R4DigitRevIndexTableGen &
digit_reverse. R4DigitRevIndexTableGen creates index tables that are used by the
digit_reverse.

w points to a floating-point array of length (3/2)*n which holds the (3/4)*n complex twiddle
factors. w can be created with the radix 4 twiddle generation function found in the FFT
support: tw_genr4fft.

n is the length of the FFT in complex samples. n must be a power of 4.

Signal Processing Examples Using the TMS320C67x Digital Signal Processing Library (DSPLIB)

SPRA947A

3.3.3

DSPF_sp_fftSPxSP Cache Optimized Mixed Radix FFT with digit reversal

void DSPF_sp_fftSPxSP(int n, float* x, float* w, float* y, unsigned char
brev[], int n_min, int offset, int n_max)

n is the length of the FFT in complex samples. nmust be a power of 2 and greater than or
equal to 8 and less than or equal to 8192.

x points to a floating-point array of length 2*n which holds n complex input samples. x must
be double-word aligned.

w points to a floating-point array of length 2*n which holds n complex twiddle factors. w can
be created with the twiddle generation function found in FFT support: tw_genSPxSPfft.

y points to a floating-point array of length 2*n which holds n complex output samples. y must
be double-word aligned.

brev is a 64-entry bit-reverse data table. The values for this table can be found in the FFT
support file brev_table.h

n_min is the smallest FFT butterfly used in computation.

offset is the index of complex FFT samples from the start of the main FFT.

n_max size of the main FFT in complex samples.

The DSPF_sp_fftSPxSP routine has been modified to allow for higher cache efficiency. The
routine can be called in a single-pass or multi-pass fashion. As single-pass, the routine behaves
like other DSPLIB FFT routines: if the total data size accessed by the routine fits into the L1D,
then the single-pass use is most efficient. The total data size accessed for an N-point FFT is N x
2 complex parts x 4 bytes per floating point input value plus the same amount for the twiddle
factor array: 16XN bytes. The L1D capacity for the C671x device is 4 Kbytes. If N less than or
equal to 256, the single pass is the best choice. If N is greater than 256, then the multi-pass
implementation would be the best choice. For more details on cache, see the TMS320 DSP
Cache Users Guide (SPRU656).
3.3.3.1

DSPF_sp_fftSPxSP Single-Pass Implementation

The single-pass implementation is straight forward.

n = n_max

x, w, and y all point to their start of the arrays

n_min = the radix of the FFT

offset =0

For example when N =256 and radix = 4:

DSPF_sp_fftSPxSP(N, &x[0], &w[0], y, brev, radix, 0, N);

3.3.3.2

DSPF_sp_fftSPxSP Multi-Pass Implementation

The multi-pass implementation requires multiple calls of the same function. The goal of this
implementation is to break up a large FFT into several FFTs that are small enough to fit into the
L1D (N <= 256). For example, a 1024 length FFT would be broken up into 4 256 length
sub-FFTs. Similarly, a 2K length FFT would be broken into 16 128 length sub-FFTs. (By nature
of the function, there must be a power of 4 sub-FFTs, i.e., 4 or 16.)

Signal Processing Examples Using the TMS320C67x Digital Signal Processing Library (DSPLIB)

13

SPRA947A

The multi-pass implementation requires two stages. The first stage only has one function call:

n = n_max

x, w, and y all point to start of the array

n_min = n divided by the number of sub-FFTs

offset =0

The second stage computes the individual sub-FFTs. The second stage has 1 call for each
sub-FFT:

n = N divided by the number of sub-FFTs. This value corresponds to the length of the
sub-FFT.

x is offset to point to the start of the sub-FFT array.

w points to the twiddle factors for the second stage. This pointer is the same for each of the
sub-FFTs (see example).

y points to the start of the array

n_min is the radix of the sub-FFT.

offset is the integer offset that corresponds to the start of the sub-FFT in the input data set.

N=N

For example, when N = 1024 and radix= 4, the multi-pass implementation requires four passes:
/* stage one */
DSPF_sp_fftSPxSP(N,

&x[0],

&w[0],

y, brev, N/4,

0,

N);

/* stage two */
DSPF_sp_fftSPxSP(N/4,&x[2*3*N/4], &w[2*3*N/4], y, brev, radix, N*3/4, N);
DSPF_sp_fftSPxSP(N/4,&x[2*2*N/4], &w[2*3*N/4], y, brev, radix, N*2/4, N);
DSPF_sp_fftSPxSP(N/4,&x[2*1*N/4], &w[2*3*N/4], y, brev, radix, N*1/4, N);
DSPF_sp_fftSPxSP(N/4,&x[0],

&w[2*3*N/4], y, brev, radix, 0,

N);

Also, when N = 2048 and radix = 2, the multi-pass implementation requires 16 passes:
/* stage one */
DSPF_sp_fftSPxSP(N, &x[0], &w[0], y, brev, N/16, 0, N);
/* stage two */
for(i=0;i<16;i++)
{
DSPF_sp_fftSPxSP(N/16, &x[2*(15i)*N/16], &w[2*N*15/16], &y[0], brev, radix,
15i)*N/16, N);
}

14

Signal Processing Examples Using the TMS320C67x Digital Signal Processing Library (DSPLIB)

SPRA947A

3.3.4

FFT Example
This example demonstrates the use of the C67x DSPLIB FFT filtering capabilities.
Input data to the FFT is generated in floating point format as follows:
for(i=0; i<N; i++)
{
/* real part */
x[2 * i] = (sin(2 * PI * F1 * i / N) + sin(2 * PI * F2 * i / N));

/* img part */
x[2 * i + 1] = 0;
}

Where F1 and F2 are the input frequencies and N is the length of the FFT. Figure 9 shows the
real part input of the FFT where F1 = 10, F2 = 40, and N = 512. Figure 10 shows the magnitude
of the output.

Figure 9. 512-Point FFT Input

Signal Processing Examples Using the TMS320C67x Digital Signal Processing Library (DSPLIB)

15

SPRA947A

Figure 10. Magnitude of the Output

3.3.5

Performance Analysis
Table 7 shows the performance of a 1024 point FFT using the single-pass implementation of
DSPF_sp_fftSPxSP.
Table 7. Single-Pass FFT Benchmarks for 1024 Point input
Number of Cycles
Functions

Formula

Observed (with data in L2 SRAM)

DSPF_sp_fftSPxSP

14464=3*ceil(log4(N)1)*N + 21*
ceil(log4(N)1) + 2*N + 44

28444

Clearly, 28,444 cycles is significantly more than the cycle count formula of 14,464. As explained
above, the single-pass implementation creates a significant amount of cache trashing. Table 8
shows the cache advantages of the multi-pass implementation by comparing the cache misses
of the single-pass and multi-pass implementations.
Table 8. Single-Pass vs. Multi-Pass FFT Benchmarks
First Call

Second Call Third Call

Fourth Call

Fifth Call

Total Observed

Single-Pass
DSPF_sp_fftSPxSP

28444

L1D Miss Cycles

13870

13870

L1P Miss Cycles

122

122

Multi-Pass
DSPF_sp_fftSPxSP

16

24976

L1D Miss Cycles

5724

1107

1013

1019

1081

9944

L1P Miss Cycles

93

30

138

Signal Processing Examples Using the TMS320C67x Digital Signal Processing Library (DSPLIB)

SPRA947A

The total cycle count for the multi-pass implementation is significantly less than the single-pass
implementation. The multi-pass implementation allows less L1D miss cycles in each of the
sub-FFTs because all the data used by each 256 length sub-FFT fits into the 4 Kbyte cache.

References
1.
2.
3.
4.
5.

TMS320C67x DSP Library Programmers Reference (SPRU657)

TMS320C6000 Peripherals Reference Guide (SPRU190)
TMS320C621x/C671x DSP Two-Level Internal Memory Reference Guide (SPRU609)
TMS320 DSP Cache Users Guide (SPRU656)
John G. Proakis and Dimitris G. Manolakis, Digital Signal Processing, Principles Algorithms,
and Applications, Prentice Hall, Third Edition, 1996.
6. Emmanuel C. Ifeachor and Barrie W. Jervis, Digital Signal Processing, A Practical Approach,
Prentice Hall, Second Edition, 2002.

Signal Processing Examples Using the TMS320C67x Digital Signal Processing Library (DSPLIB)

17

IMPORTANT NOTICE
Texas Instruments Incorporated and its subsidiaries (TI) reserve the right to make corrections, modifications, enhancements, improvements,
and other changes to its products and services at any time and to discontinue any product or service without notice. Customers should
obtain the latest relevant information before placing orders and should verify that such information is current and complete. All products are
sold subject to TIs terms and conditions of sale supplied at the time of order acknowledgment.
TI warrants performance of its hardware products to the specifications applicable at the time of sale in accordance with TIs standard
warranty. Testing and other quality control techniques are used to the extent TI deems necessary to support this warranty. Except where
mandated by government requirements, testing of all parameters of each product is not necessarily performed.
TI assumes no liability for applications assistance or customer product design. Customers are responsible for their products and
applications using TI components. To minimize the risks associated with customer products and applications, customers should provide
adequate design and operating safeguards.
TI does not warrant or represent that any license, either express or implied, is granted under any TI patent right, copyright, mask work right,
or other TI intellectual property right relating to any combination, machine, or process in which TI products or services are used. Information
published by TI regarding third-party products or services does not constitute a license from TI to use such products or services or a
warranty or endorsement thereof. Use of such information may require a license from a third party under the patents or other intellectual
property of the third party, or a license from TI under the patents or other intellectual property of TI.
Reproduction of TI information in TI data books or data sheets is permissible only if reproduction is without alteration and is accompanied
by all associated warranties, conditions, limitations, and notices. Reproduction of this information with alteration is an unfair and deceptive
business practice. TI is not responsible or liable for such altered documentation. Information of third parties may be subject to additional
restrictions.
Resale of TI products or services with statements different from or beyond the parameters stated by TI for that product or service voids all
express and any implied warranties for the associated TI product or service and is an unfair and deceptive business practice. TI is not
responsible or liable for any such statements.
TI products are not authorized for use in safety-critical applications (such as life support) where a failure of the TI product would reasonably
be expected to cause severe personal injury or death, unless officers of the parties have executed an agreement specifically governing
such use. Buyers represent that they have all necessary expertise in the safety and regulatory ramifications of their applications, and
acknowledge and agree that they are solely responsible for all legal, regulatory and safety-related requirements concerning their products
and any use of TI products in such safety-critical applications, notwithstanding any applications-related information or support that may be
provided by TI. Further, Buyers must fully indemnify TI and its representatives against any damages arising out of the use of TI products in
such safety-critical applications.
TI products are neither designed nor intended for use in military/aerospace applications or environments unless the TI products are
specifically designated by TI as military-grade or "enhanced plastic." Only products designated by TI as military-grade meet military
specifications. Buyers acknowledge and agree that any such use of TI products which TI has not designated as military-grade is solely at
the Buyer's risk, and that they are solely responsible for compliance with all legal and regulatory requirements in connection with such use.
TI products are neither designed nor intended for use in automotive applications or environments unless the specific TI products are
designated by TI as compliant with ISO/TS 16949 requirements. Buyers acknowledge and agree that, if they use any non-designated
products in automotive applications, TI will not be responsible for any failure to meet such requirements.
Following are URLs where you can obtain information on other Texas Instruments products and application solutions:
Products
Amplifiers
Data Converters
DLP Products
DSP
Clocks and Timers
Interface
Logic
Power Mgmt
Microcontrollers
RFID
RF/IF and ZigBee Solutions

amplifier.ti.com
dataconverter.ti.com
www.dlp.com
dsp.ti.com
www.ti.com/clocks
interface.ti.com
logic.ti.com
power.ti.com
microcontroller.ti.com
www.ti-rfid.com
www.ti.com/lprf

Applications
Audio
Automotive
Broadband
Digital Control
Medical
Military
Optical Networking
Security
Telephony
Video & Imaging
Wireless

www.ti.com/audio
www.ti.com/automotive
www.ti.com/broadband
www.ti.com/digitalcontrol
www.ti.com/medical
www.ti.com/military
www.ti.com/opticalnetwork
www.ti.com/security
www.ti.com/telephony
www.ti.com/video
www.ti.com/wireless

Mailing Address: Texas Instruments, Post Office Box 655303, Dallas, Texas 75265
Copyright 2009, Texas Instruments Incorporated