Application Report

SPRABB6B– June 2010 – Revised January 2013

FFT Implementation on the TMS320VC5505,

TMS320C5505, and TMS320C5515 DSPs
Mark McKeown

ABSTRACT
The Fast Fourier Transform (FFT) is an efficient means for computing the Discrete Fourier Transform
(DFT). It is one of the most widely used computational elements in Digital Signal Processing (DSP)
applications. This DSP is ideally suited for such applications. They include an FFT hardware accelerator
(HWAFFT) that is tightly coupled with the CPU, allowing high FFT processing performance at very low
power. This application report describes FFT computation on the DSP and covers the following topics:
• Basics of DFT and FFT
• DSP Overview Including the FFT Accelerator
• HWAFFT Description
• HWAFFT Software Interface
• Simple Example to Illustrate the Use of the FFT Accelerator
• FFT Benchmarks
• Description of Open Source FFT Example Software
• Computation of Large (Greater Than 1024-point) FFTs
Project collateral and source code discussed in this application report can be downloaded from the
following URL: http://www.ti.com/lit/zip/SPRABB6.

Contents
1 Basics of DFT and FFT .......................................................................................................................... 2
2 DSP Overview Including the FFT Accelerator ........................................................................................ 6
3 FFT Hardware Accelerator Description .................................................................................................. 8
4 HWAFFT Software Interface ................................................................................................................ 11
5 Simple Example to Illustrate the Use of the FFT Accelerator ............................................................... 17
6 FFT Benchmarks .................................................................................................................................. 20
7 Description of Open Source FFT Example Software............................................................................ 21
8 Computation of Large (Greater Than 1024-Point) FFTs ...................................................................... 22
9 Appendix A Methods for Aligning the Bit-Reverse Destination Vector ............................................................... 25
Appendix A Revision History...................................................................................................................................... 27

List of Figures
1 DIT Radix 2 Butterfly .............................................................................................................................. 3
2 DIT Radix 2 8-point FFT ......................................................................................................................... 4
3 Graphical FFT Computation ................................................................................................................... 5
4 Block Diagram ........................................................................................................................................ 7
5 Bit Reversed Input Buffer ..................................................................................................................... 13
6 Graphing the Real Part of the FFT Result in CCS4 ............................................................................. 19
7 Graphing the Imaginary Part of the FFT Result in CCS4 ..................................................................... 19
8 FFT Filter Demo Block Diagram ........................................................................................................... 21

List of Tables

SPRABB6B– June 2010 – Revised January 2013 FFT Implementation on the TMS320VC5505, TMS320C5505, and 1
Submit Documentation Feedback TMS320C5515 DSPs
Copyright © 2010–2013, Texas Instruments Incorporated
Basics of DFT and FFT www.ti.com

1 Computational Complexity of Direct DFT Computation versus Radix-2 FFT ....................................................... 4

2 Available HWAFFT Routines .............................................................................................................................. 15
3 FFT Performance on HWAFFT vs CPU (Vcore = 1.05 V, PLL = 60 MHz) ......................................................... 20
4 FFT Performance on HWAFFT vs CPU (Vcore = 1.3 V, PLL = 100 MHz) ......................................................... 20
5 Revision History ................................................................................................................................................. 27

1 Basics of DFT and FFT

The DFT takes an N-point vector of complex data sampled in time and transforms it to an N-point vector of
complex data that represents the input signal in the frequency domain. A discussion on the DFT and FFT
is provided as background material before discussing the HWAFFT implementation.
The DFT of the input sequence x(n), n = 0, 1, …, N-1 is defined as
1 N -1
X (k ) = x n W nk k to N
Ι ( ) , =0 -1
Nn=0 N (1)
Where WN, the twiddle factor, is defined as
- j 2ν / N
W =e , k = 0 to N - 1
N (2)
The FFT is a class of efficient DFT implementations that produce results identical to the DFT in far fewer
cycles. The Cooley-Tukey algorithm is a widely used FFT algorithm that exploits a divide-and-conquer
approach to recursively decompose the DFT computation into smaller and smaller DFT computations until
the simplest computation remains. One subset of this algorithm called Radix-2 Decimation-in-Time (DIT)
breaks each DFT computation into the combination of two DFTs, one for even-indexed inputs and another
for odd-indexed inputs. The decomposition continues until a DFT of just two inputs remains. The 2-point
DFT is called a butterfly, and it is the simplest computational kernel of Radix-2 FFT algorithms.

1.1 Radix-2 Decimation in Time Equations

The Radix-2 DIT decomposition can be seen by manipulating the DFT equation (Equation 1):
Split x(n) into even and odd indices:
1 N / 2 -1 1 N / 2 -1 (2n + 1)k
X (k ) = x(2n)W 2nk x n W k to N
Ι + Ι (2 + 1) , =0 -1
Nn=0 N N n =0 N (3)
Wk
Factor N from the odd indexed summation:
W k N / 2 -1
N / 2 -1
1
X (k ) = Ι, x(2n)W 2nk + N Ι, x(2n + 1)W 2nk , k = 0 to N - 1
N n =0 N Nn=0 N (4)
Only twiddle factors from 0 to N/2 are needed:
- j 2ΙΤ / N 2nk - j 2ΙΤ /(N / 2) nk k + N /2
W 2nk = (e ) = (e ) = W nk and W = - W k , k = 0 to N / 2 - 1
N N /2 N N (5)
This results in:
W k N / 2 -1
1N / 2 -1 nk
X (k ) = Ι, x(2n)W + N Ι, x(2n + 1)W nk , k = 0 to N - 1
Nn=0 N /2 N n =0 N/2 (6)
X (k ) X (k )
Define even and odd such that:
1 N / 2 -1 nk
X (k ) = x(2n)W , k = 0 to N - 1
even (N / 2) n = 0 N/2 (7)
and
1 N / 2 -1
X (k ) = x(2n + 1)W nk , k = 0 to N - 1
odd (N / 2) n = 0 N/2 (8)

2 FFT Implementation on the TMS320VC5505, TMS320C5505, and SPRABB6B– June 2010 – Revised January 2013
TMS320C5515 DSPs Submit Documentation Feedback
Copyright © 2010–2013, Texas Instruments Incorporated
www.ti.com Basics of DFT and FFT

Finally, Equation 6 is rewritten as:

1
X (k ) = (X (k ) + W k X (k )), k = 0 to N / 2 - 1
2 even N odd (9)
and
1
X (k + N / 2) = (X (k ) - W k X (k )), k = 0 to N / 2 - 1
2 even N odd (10)
Equation 9 and Equation 10 show that the N-point DFT can be divided into two smaller N/2-point DFTs.
Each smaller DFT is then further divided into smaller DFTs until N = 2. The pair of equations that makeup
the 2-point DFT is called the Radix2 DIT Butterfly (see Section 1.2). The DIT Butterfly is the core
calculation of the FFT and consists of just one complex multiplication and two complex additions.

1.2 Radix-2 DIT Butterfly

The Radix-2 Butterfly is illustrated in Figure 1. In each butterfly structure, two complex inputs P and Q are
operated upon and become complex outputs P’ and Q’. Complex multiplication is performed on Q and the
twiddle factor, then the product is added to and subtracted from input P to form outputs P’ and Q’. The
Wk
exponent of the twiddle factor N is dependent on the stage and group of its butterfly. The butterfly is
usually represented by its flow graph (Figure 1), which looks like a butterfly.

k
P + P’ = P + Q∗W N
+

Q + k
Q’ = P – Q∗W N
W k –
N

Figure 1. DIT Radix 2 Butterfly

The mathematical meaning of this butterfly is shown below with separate equations for real and imaginary
parts:

Complex Real Part Imaginary Part

P’ = P + Q * W Pr’ = Pr + (Qr * Wr - Qi * Wi) Pi’ = Pi + (Qr * Wi + Qi * Wr)
Q’ = P - Q * W Qr’ = Pr - (Qr * Wr - Qi * Wi) Qi’ = Pi - (Qr * Wi + Qi * Wr)

The flow graph in Figure 2 shows the interconnected butterflies of an 8-point Radix-2 DIT FFT. Notice that
the inputs to the FFT are indexed in bit-reversed order (0, 4, 2, 6, 1, 5, 3, 7) and the outputs are indexed
in sequential order (0, 1, 2, 3, 4, 5, 6, 7). Computation of a Radix-2 DIT FFT requires the input vector to be
in bit-reversed order, and generates an output vector in sequential order. This bit-reversal is further
explained in Section 4.3, Bit-Reverse Function.

SPRABB6B– June 2010 – Revised January 2013 FFT Implementation on the TMS320VC5505, TMS320C5505, and 3
Submit Documentation Feedback TMS320C5515 DSPs
Copyright © 2010–2013, Texas Instruments Incorporated
Basics of DFT and FFT www.ti.com

Figure 2. DIT Radix 2 8-point FFT

1.3 Computational Complexity

The Radix-2 DIT FFT requires log2(N) stages, N/2 * log2(N) complex multiplications, and N * log2(N)
complex additions. In contrast, the direct computation of X(k) from the DFT equation (Equation 1) requires
N2 complex multiplications and (N2 – N) complex additions. Table 1 compares the computational
complexity for direct DFT versus Radix-2 FFT computations for typical FFT lengths.

Table 1. Computational Complexity of Direct DFT Computation versus Radix-2 FFT

FFT Length Direct DFT Computation Radix-2 FFT
Complex Complex Additions Complex Complex Additions
Multiplications Multiplications
128 16,384 16,256 448 896
256 65,536 65,280 1,024 2,048
512 262,144 264,632 2,304 4,608
1024 1,048,576 1,047,552 5,120 10,240

Table 1 clearly shows significant reduction in computational complexity with the Radix-2 FFT, especially
for large N. This substantial decrease in computational complexity of the FFT has allowed DSPs to
efficiently compute the DFT in reasonable time. For its substantial efficiency improvement over direct
computation, the HWAFFT coprocessor in the DSP implements the Radix-2 FFT algorithm.

4 FFT Implementation on the TMS320VC5505, TMS320C5505, and SPRABB6B– June 2010 – Revised January 2013
TMS320C5515 DSPs Submit Documentation Feedback
Copyright © 2010–2013, Texas Instruments Incorporated
www.ti.com Basics of DFT and FFT

1.4 FFT Graphs

Figure 3 is a graphical example of the FFT computation. These results were obtained by using the
HWAFFT coprocessor on the DSP. On the left half is the complex time domain signal (real part on top,
imaginary part on bottom). On the right half is the complex frequency domain signal produced by the FFT
computation (real part on top, imaginary part on bottom). In this example, two sinusoidal tones are present
in the time domain. The time domain signal is 1024 points and contains only real data (the imaginary part
is all zeros). The two sinusoids are represented in the frequency domain as impulses located at the
frequency bins that correspond to the two sinusoidal frequencies. The frequency domain signal is also
1024 points and contains both real parts (top right) and imaginary parts (bottom right).

Real (Time Domain Signal) Real (Frequency Domain Signal)

FFT
IFFT
Imaginary (Time Domain Signal) Imaginary (Frequency Domain Signal)

Figure 3. Graphical FFT Computation

SPRABB6B– June 2010 – Revised January 2013 FFT Implementation on the TMS320VC5505, TMS320C5505, and 5
Submit Documentation Feedback TMS320C5515 DSPs
Copyright © 2010–2013, Texas Instruments Incorporated
DSP Overview Including the FFT Accelerator www.ti.com

2 DSP Overview Including the FFT Accelerator

This DSP is a member of TI's TMS320C5000™ fixed-point DSP product family and is designed for low-
power applications. With an active mode power consumption of less than 0.15 mW/MHz and a standby
mode power consumption of less than 0.15 mW, these DSPs are optimized for applications characterized
by sophisticated processing and portable form factors that require low power and longer battery life.
Examples of such applications include portable voice/audio devices, noise cancellation headphones,
software-defined radio, musical instruments, medical monitoring devices, wireless microphones,
biometrics, industrial instruments, telephony, and audio cards.
Figure 4 shows an overview of the DSP consisting of the following primary components:
• Dual MAC, C55x CPU
1.05 V @ 60 MHz (VC5505, C5505, C5515) and 75 MHz (C5505 and C5515)
1.3 V @ 100 MHz (VC5505, C5505, C5515) and 120 MHz (C5505 and C5515)
1.4 V @ 150 MHz (C5505)
• On-Chip memory: 320 KB RAM (64 KB DARAM, 256 KB SARAM), 128 KB ROM
• HWAFFT that supports 8- to 1024-point (powers of 2) real and complex-valued FFTs
• Four DMA controllers and external memory interface
• Power management module
• A set of I/O peripherals that includes I2C, I2S, SPI, UART, Timers, EMIF, MMC/SD, GPIO, 10-bit SAR,
LCD Controller, USB 2.0
• Three on-chip LDO Regulators (C5515), 1 on-chip LDO Regulator (VC5505, C5505)
• SDRAM/mSDRAM support (C5505 and C5515)

6 FFT Implementation on the TMS320VC5505, TMS320C5505, and SPRABB6B– June 2010 – Revised January 2013
TMS320C5515 DSPs Submit Documentation Feedback
Copyright © 2010–2013, Texas Instruments Incorporated
www.ti.com DSP Overview Including the FFT Accelerator

6 6 2 4 16

3 Timers
MMC/SD

MMC/SD

USB 2.0
slave
GPIO

GPIO

RTC
IC
IS

IS
2

2
4-Ch.
EMIF/NAND, DMA
60 C55x
TM

SDRAM/ 2
mSDRAM DSP Core

INT
Peripheral Bus
FFT

Memory
DARAM
4 64 KB SARAM
32KHz 7

JTAG
ROM 256 KB
PLL
128 KB

3 LDOs
UART

10-Bit
GPIO

SAR
LCD
SPI
IS

IS
2

4 4 4 7
8 8

8 13

Figure 4. Block Diagram

SPRABB6B– June 2010 – Revised January 2013 FFT Implementation on the TMS320VC5505, TMS320C5505, and 7
Submit Documentation Feedback TMS320C5515 DSPs
Copyright © 2010–2013, Texas Instruments Incorporated
FFT Hardware Accelerator Description www.ti.com

The C55x CPU includes a tightly coupled FFT accelerator that communicates with the C55x CPU through
the use of the coprocessor instructions. The main features of this hardware accelerator are:
• Supports 8- to 1024-point (powers of 2) complex-valued FFTs.
• Internal twiddle factor generation for optimal use of memory bandwidth and more efficient
programming.
• Basic and software-driven auto-scaling feature provides good precision versus cycle count trade-off.
• Single-stage and double-stage modes enable computation of one or two stages in one pass, and thus
better handle the odd power of two FFT widths.
• Is 4 to 6 times more energy efficient and 2.2 to 3.8 times faster than the FFT computations on the
CPU.

3 FFT Hardware Accelerator Description

The HWAFFT in the DSP is a tightly-coupled, software-controlled coprocessor designed to perform FFT
and inverse FFT (IFFT) computations on complex data vectors ranging in length from 8 to 1024 points
(powers of 2). It implements a Radix-2 DIT structure that returns the FFT or IFFT result in bit-reversed
order.

3.1 Tightly-Coupled Hardware Accelerator

The HWAFFT is tightly-coupled with the DSP core which means that it is physically located outside of the
DSP core but has access to the core’s full memory read bandwidth (busses B, C, and D), access to the
core’s internal registers and accumulators, and access to its address generation units. The HWAFFT
cannot access the data write busses or memory mapped registers (MMRs). Because the HWAFFT is seen
as part of the execution unit of the CPU, it must also comply to the core’s pipeline exceptions, and in
particular those caused by stalls and conditional execution.

3.2 Hardware Butterfly, Double-Stage and Single-Stage Mode

The core of the HWAFFT consists of a single Radix-2 DIT Butterfly implemented in hardware. This
hardware supports a double-stage mode where two FFT stages are computed a single pass. In this mode
the HWAFFT feeds the results from the first stage back into the hardware butterfly to compute the second
stage results in a single pass. This double-stage mode offers significant speed-up especially for large FFT
lengths. However, when the number of required stages is odd (FFT lengths = 8, 32, 128, or 512 points)
the final stage needs to be computed alone and, consequently, at a lower acceleration rate. For this
reason a single-stage mode is also provided.
The HWAFFT supports two stage modes:
• Double-Stage Mode – two FFT stages performed in each pass
• Single-Stage Mode – one FFT stage performed in each pass

3.3 Pipeline and Latency

The logic of the HWAFFT is pipelined to deliver maximum throughput. Complex multiplication with the
twiddle factors is performed in the first pipeline stage, and complex addition and subtraction is performed
in the second pipeline stage. Valid results appear some cycles of latency after the first data is read from
memory:
• 5 cycles of latency in single-stage mode
• 9 cycles of latency in double-stage mode
There are three states to consider during computation of a single or double stage:
• Prologue: The hardware accelerator is fed with one complex input at a time but does not output any
valid data.
• Kernel: Valid outputs appear while new inputs are received and computed upon.
• Epilogue: A few more cycles are needed to flush the pipeline and output the last butterfly results.

8 FFT Implementation on the TMS320VC5505, TMS320C5505, and SPRABB6B– June 2010 – Revised January 2013
TMS320C5515 DSPs Submit Documentation Feedback
Copyright © 2010–2013, Texas Instruments Incorporated
www.ti.com FFT Hardware Accelerator Description

Consecutive stages can be overlapped such that the first data points for the next pass are read while the
final output values of the current pass are returned. For odd-power-of-two FFT lengths, the last double-
stage pass needs to be completed before starting a final single-stage pass. Thus, the double-stage
latency is only experienced once for even-powers-of-2 FFT computations and twice for odd-powers-of-2
FFT computations. Latency has little impact on the total computation performance, and less and less so
as the FFT size increases.

3.4 Software Control

Software is required to communicate between the CPU and the HWAFFT. The CPU instruction set
architecture (ISA) includes a class of coprocessor (copr) instructions that allows the CPU to initialize, pass
data to, and execute butterfly computations on the HWAFFT. Computation of an FFT/IFFT is performed by
executing a sequence of these copr instructions.
C-callable HWAFFT functions are provided with optimized sequences of copr instructions for each
available FFT length. To conserve program memory, these functions are located in the DSP’s ROM. A
detailed explanation of the HWAFFT software interface and its application is provided in Section 4,
HWAFFT Software Interface.

NOTE: To execute the HWAFFT routines from the ROM of the DSP, the programmer must satisfy
memory allocation restrictions for the data and scratch buffers. See the device-specific errata
for an explanation of the restrictions and workarounds:
• TMS320VC5505/VC5504 Fixed-Point DSP Silicon Errata (Silicon Revision 1.4)
[literature number SPRZ281]
• TMS320C5505/C5504 Fixed-Point DSP Silicon Errata (Silicon Revision 2.0)
[literature number SPRZ310]
• TMS320C5515/C5514 Fixed-Point DSP Silicon Errata (Silicon Revision 2.0)
[literature number SPRZ308]

3.5 Twiddle Factors

To conserve memory bus bandwidth, twiddle factors are stored in a look-up-table within the HWAFFT
coprocessor. The 512 complex twiddle factors (16-bit real, 16-bit imaginary) are available for computing up
to 1024-point FFTs. Smaller FFT lengths use a decimated subset of these twiddle factors. Indexing the
twiddle table is pipelined and optimized based on the current FFT stage and group being computed. When
the IFFT computation is performed, the complex conjugates of the twiddle factors are used.

SPRABB6B– June 2010 – Revised January 2013 FFT Implementation on the TMS320VC5505, TMS320C5505, and 9
Submit Documentation Feedback TMS320C5515 DSPs
Copyright © 2010–2013, Texas Instruments Incorporated
FFT Hardware Accelerator Description www.ti.com

3.6 Scaling
FFT computation with fixed-point numbers is vulnerable to overflow, underflow, and saturation effects.
Depending on the dynamic range of the input data, some scaling may be required to avoid these effects
during the FFT computation. This scaling can be done before the FFT computation, by computing the
dynamic range of the input points and scaling them down accordingly. If the magnitude of each complex
input element is less than 1/N, where N = FFT Length, then the N-point FFT computation will not overflow.
Uniformly dividing the input vector elements by N (Pre-scaling) is equivalent to shifting each binary
number right by log2(N) bits, which introduces significant error (especially for large FFT lengths). When
this error propagates through the FFT flow graph, the output noise-to-signal ratio increases by 1 bit per
stage or log2(N) bits in total. Overflow will not occur if each input’s magnitude is less than 1/N.
Alternatively, a simple divide-by-2 and round scaling after each butterfly offers a good trade-off between
precision and overflow protection, while minimizing computation cycles. Because the error introduced by
early FFT stages is also scaled after each butterfly, the output noise-to-signal ratio increases by just ½ bit
per stage or ½ * log2(N) bits in total. Overflow is avoided if each input’s magnitude is less than 1.
The HWAFFT supports two scale modes:
• NOSCALE
– Scaling logic disabled
– Vulnerable to overflow
– Output dynamic range grows with each stage
– No overflow if input magnitudes < 1/N
• SCALE
– Scales each butterfly output by 1/2
– No overflow if input magnitudes < 1

10 FFT Implementation on the TMS320VC5505, TMS320C5505, and SPRABB6B– June 2010 – Revised January 2013
TMS320C5515 DSPs Submit Documentation Feedback
Copyright © 2010–2013, Texas Instruments Incorporated
www.ti.com HWAFFT Software Interface

4 HWAFFT Software Interface

The software interface to the HWAFFT is handled through a set of coprocessor instructions that, when
decoded by the coprocessor, perform initialization, load/store, and execution operations on the HWAFFT
coprocessor. C-callable functions are provided that contain the necessary sequences of coprocessor
instructions for performing FFT/ IFFT computations in the range of 8 to 1024 points (by powers of 2).
Additionally, an optimized out-of-place bit-reversal function is provided to perform the complex vector bit-
reversal required by Radix-2 FFT computations. These functions are defined in the hwafft.asm source
code file. Additionally, to conserve on-chip RAM these functions have been placed in the on-chip ROM of
the DSP. See Section 4.5, Project Configuration for Calling Function s from ROM, for steps to configure
your project to call the HWAFFT functions from ROM.

4.1 Data Types

The input and output vectors of the HWAFFT contain complex numbers. Each real and imaginary part is
represented by a two’s complement, 16-bit fixed-point number. The most significant bit holds the number’s
sign value, and the remaining 15 are fraction bits (S16Q15 format). The range of each number is [-1, 1 –
(1/2)15]. Real and imaginary parts appear in an interleaved order within each vector:
Int16 CMPLX_Vec16[2*N] = …(N = FFT Length)
Real[0] Imag[0] Real[1] Imag[1] Real[2] Imag[2]
Bit15,. ................ ,Bit0 Bit15,. ................ ,Bit0 Bit15,. ................ ,Bit0 Bit15,. ................ ,Bit0 Bit15,. ................ ,Bit0 Bit15,. ................ ,Bit0

The HWAFFT functions use an Int32 pointer to reference these complex vectors. Therefore, each 32-bit
element contains the 16-bit real part in the most significant 16 bits, and the 16-bit imaginary part in the
least significant 16 bits.
Int32 CMPLX_Vec32[N] = …(N = FFT Length)
Real[0] Imag[0] Real[1] Imag[1] Real[2] Imag[2]
Bit31,.................., Bit16, Bit15,. ................. , Bit0 Bit31,.................., Bit16, Bit15,. ................ , Bit0 Bit31,.................., Bit16, Bit15,. ................. , Bit0

To extract the real and imaginary parts from the complex vector, it is necessary to mask and shift each 32-
bit element into its 16-bit real and imaginary parts:
Uint16 Real_Part = CMPLX_Vec32[i] >> 16;
Uint16 Imaginary_Part = CMPLX_Vec32[i] & 0x0000FFFF;

4.2 HWAFFT Functions

C-Callable HWAFFT Functions are provided for computing FFT/IFFT transforms on the HWAFFT
coprocessor. These functions contain optimized sequences of coprocessor instructions for computing
scaled or unscaled 8-, 16-, 32-, 64-, 128-, 256-, 512-, and 1024-point FFT/IFFTs. Additionally, an
optimized out-of-place bit-reversal function is provided to bit-reverse the input vector before supplying it to
the HWAFFT. Computation of a Radix-2 DIT FFT requires the input vector to be in bit-reversed order, and
generates an output vector in sequential order.

4.2.1 HWAFFT Naming and Format

NOTE: To execute the HWAFFT routines from the ROM of the DSP, the programmer must satisfy
memory allocation restrictions for the data and scratch buffers. See the device-specific errata
for an explanation of the restrictions and workarounds:
• TMS320VC5505/VC5504 Fixed-Point DSP Silicon Errata (Silicon Revision 1.4)
[literature number SPRZ281]
• TMS320C5505/C5504 Fixed-Point DSP Silicon Errata (Silicon Revision 2.0)
[literature number SPRZ310]
• TMS320C5515/C5514 Fixed-Point DSP Silicon Errata (Silicon Revision 2.0)
[literature number SPRZ308]

SPRABB6B– June 2010 – Revised January 2013 FFT Implementation on the TMS320VC5505, TMS320C5505, and 11
Submit Documentation Feedback TMS320C5515 DSPs
Copyright © 2010–2013, Texas Instruments Incorporated
HWAFFT Software Interface www.ti.com

The HWAFFT functions are named hwafft_Npts, where N is the FFT length. For example, hwafft_32pts is
the name of the function for performing 32-point FFT and IFFT operations. The structure of the HWAFFT
functions is:
Uint16 hwafft_Npts( Performs N-point complex FFT/IFFT, where N = {8, 16, 32, 64, 128, 256, 512,
1024}
Int32 *data, Input/output – complex vector
Int32 *scratch, Intermediate/output – complex vector
Uint16 Flag determines whether FFT or IFFT performed, (0 = FFT, 1 = IFFT)
fft_flag,
Uint16 Flag determines whether butterfly output divided by 2 (0 = Scale, 1 = No Scale)
scale_flag
); Return value Flag determines whether output in data or scratch vector (0 = data, 1 = scratch)

4.2.2 HWAFFT Parameters

The following describe the parameters for the HWAFFT functions.
Int32 *data
This is the input vector to the HWAFFT. It contains complex data elements (real part in most significant 16
bits, imaginary part in least significant 16 bits). After the HWAFFT function completes, the result will either
be stored in this data vector or in the scratch vector, depending on the status of the return value. The
return value is Boolean where 0 indicates that the result is stored in this data vector, and 1 indicates the
scratch vector. The data and scratch vectors must reside in separate blocks of RAM (DARAM or SARAM)
to maximize memory bandwidth.
#pragma DATA_SECTION(data_buf, "data_buf");
//Static Allocation to Section: "data_buf : > DARAM" in Linker CMD File
Int32 data_buf[N = FFT Length];
Int32 *data = data_buf;
Int32 *data:

The *data parameter is a complex input vector to HWAFFT. It contains the output vector if Return Value =
0 = OUT_SEL_DATA. There is a strict address alignment requirement if *data is shared with a bit-reverse
destination vector (recommended). See Section 4.3.1, Bit Reverse Destination Vector Alignment
Requirement.
Int32 *scratch
This is the scratch vector used by the HWAFFT to store intermediate results between FFT stages. It
contains complex data elements (real part in most significant 16 bits, imaginary part in least significant 16
bits). After the HWAFFT function completes the result will either be stored in the data vector or in this
scratch vector, depending on the status of the return value. The return value is Boolean, where 0 indicates
that the result is stored in the data vector, and 1 indicates this scratch vector. The data and scratch
vectors must reside in separate blocks of RAM (DARAM or SARAM) to maximize memory bandwidth.
#pragma DATA_SECTION(scratch_buf, "scratch_buf");
//Static Allocation to Section: "scratch_buf : > DARAM" in Linker CMD File
Int32 scratch_buf[N = FFT Length];
Int32 *scratch = scratch_buf;
Int32 *scratch:

The *scratch parameter is a complex scratchpad vector to HWAFFT. It contains the output vector if Return
Value = 1 = OUT_SEL_SCRATCH.
Uint16 fft_flag
The FFT/IFFT selection is controlled by setting the fft_flag to 0 for FFT and 1 for Inverse FFT.
#define FFT_FLAG ( 0 ) /* HWAFFT to perform FFT */
#define IFFT_FLAG ( 1 ) /* HWAFFT to perform IFFT */
Uint16 fft_flag:

12 FFT Implementation on the TMS320VC5505, TMS320C5505, and SPRABB6B– June 2010 – Revised January 2013
TMS320C5515 DSPs Submit Documentation Feedback
Copyright © 2010–2013, Texas Instruments Incorporated
www.ti.com HWAFFT Software Interface

fft_flag = FFT_FLAG: FFT Performed

fft_flag = IFFT_FLAG: Inverse FFT Performed

Uint16 scale_flag
The automatic scaling (divide each butterfly output by 2) feature is controlled by setting the scale_flag to 0
to enable scaling and 1 to disable scaling.
#define SCALE_FLAG ( 0 ) /* HWAFFT to scale butterfly output */
#define NOSCALE_FLAG ( 1 ) /* HWAFFT not to scale butterfly output */
Uint16 scale_flag:

scale_flag = SCALE_FLAG: Divide by 2 scaling is performed at the output of each FFT Butterfly.
scale_flag = NOSCALE_FLAG: No scaling is performed, overflow may occur if the input dynamic is
too high.

Uint16 <Return Value>:

This is the Uint16 return value of the HWAFFT functions. After the HWAFFT function completes, the result
will either be stored in the data vector or in the scratch vector, depending on the status of this return
value. The return value is Boolean where 0 indicates that the result is stored in the data vector, and 1
indicates this scratch vector. The program must check the status of the Return Value to determine the
location of the FFT/IFFT result.
#define OUT_SEL_DATA ( 0 ) /* indicates HWAFFT output located in input data vector */
#define OUT_SEL_SCRATCH ( 1 ) /* indicates HWAFFT output located in scratch vector */
Uint16 <Return Value>:

Return Value = OUT_SEL_DATA: FFT/IFFT result located in the data vector

Return Value = OUT_SEL_SCRATCH: FFT/IFFT result located in the scratch vector

4.3 Bit Reverse Function

Before computing the FFT/IFFT on the HWAFFT, the input buffer must be bit-reversed to facilitate a
Radix-2 DIT computation. This function contains optimized assembly that executes on the CPU to
rearrange the Int32 elements of the input vector by placing each element in the destination vector at the
index that corresponds to the bit-reversal of its current index. For example, in an 8-element vector, the
index of the third element is 011 in binary, then the bit-reversed index is 110 in binary or 6 in decimal, so
the third element of the input vector is copied to the sixth element of the bit-reversal destination vector.
Int32 data[8]

Data[0] Data[1] Data[2] Data[3] Data[4] Data[5] Data[6] Data[7]

Index = 000 001 010 011 100 101 110 111

Int32 data_br[8]
Data[0] Data[1] Data[2] Data[3] Data[4] Data[5] Data[6] Data[7]
Index = 000 100 010 110 001 101 011 111

Figure 5. Bit Reversed Input Buffer

SPRABB6B– June 2010 – Revised January 2013 FFT Implementation on the TMS320VC5505, TMS320C5505, and 13
Submit Documentation Feedback TMS320C5515 DSPs
Copyright © 2010–2013, Texas Instruments Incorporated
HWAFFT Software Interface www.ti.com

4.3.1 Bit Reverse Destination Vector Alignment Requirement

Strict requirements are placed on the address of the bit-reversal destination buffer. This buffer must be
aligned in RAM such that log2(4 * N) zeros appear in the least significant bits of the byte address (8 bits),
where N is the FFT Length. For example, a 1024-point FFT needs to bit-reverse 1024 complex array
elements (32-bit elements). The address for the bit-reversed buffer needs to have 12 zeros in the least
significant bits of its byte address (log2(4 * 1024) = 12). Since the word address (16 bits) is the byte
address shifted right one bit, the word address requires 11 zeros in the least significant bits. This bit-
reverse is considered out-of-place because the inputs and outputs are stored in different vectors. In-place
bit-reversal is not supported by this function. There are no alignment requirements for the bit-reverse
source vector.

4.3.2 Bit Reverse Format and Parameters

The structure of the HWAFFT functions is:

void hwafft_br( Performs out-of-place bit-reversal on 32-bit data vector

Int32 *data, Input – 32-bit data vector
Int32 Output – bit-reversed data vector
*data_br,
Uint16 Length of complex data vector
data_len,
);

The parameters for the hwafft_br function are:

Int32 *data
This is the input vector to the bit reverse function. It contains complex data elements (real part in most
significant 16 bits, imaginary part in least significant 16 bits). There are no specific alignment requirements
for this vector.
#pragma DATA_SECTION(data_buf, "data_buf");
//Static Allocation to Section: "data_buf : > DARAM" in Linker CMD File
Int32 data_buf[N = FFT Length];
Int32 *data = data_buf;

Int32 *data_br
This is the destination vector of the bit-reverse function. It contains complex data elements (real part in
most significant 16 bits, imaginary part in least significant 16 bits). A strict alignment requirement is placed
on this destination vector of the bit-reverse function: This buffer must be aligned in RAM such that log2(4 *
N) zeros appear in the least significant bits of the byte address (8 bits), where N is the FFT Length. See
Section 9, Appendix A Methods for Aligning the Bit-Reverse Destination Vector, for ways to force the
linker to enforce this alignment requirement.
#define ALIGNMENT 2*N // ALIGNS data_br_buf to an address with log2(4*N) zeros in the
// least significant bits of the byte address
#pragma DATA_SECTION(data_br _buf, "data_br_buf");
// Allocation to Section: "data_br _buf : > DARAM" in Linker CMD File
#pragma DATA_ALIGN (data_br_buf, ALIGNMENT);
Int32 data_br_buf[N = FFT Length];
Int32 * data_br = data_br_buf;
Int32 *data_br:

Strict address alignment requirement: This buffer must be aligned in RAM such that (log2(4 * N) zeros
appear in the least significant bits of the byte address (8 bits), where N is the FFT Length. See Section 9
for ways to force the linker to enforce this alignment requirement.

14 FFT Implementation on the TMS320VC5505, TMS320C5505, and SPRABB6B– June 2010 – Revised January 2013
TMS320C5515 DSPs Submit Documentation Feedback
Copyright © 2010–2013, Texas Instruments Incorporated
www.ti.com HWAFFT Software Interface

Uint16 *data_len
This Uint16 parameter indicates the length of the data and data_br vectors.
Uint16 data_len:

The data_len parameter indicates the length of the Int32 vector (FFT Length). Valid lengths include
powers of two: {8, 16, 32, 64, 128, 256, 512, 1024}.

4.4 Function Descriptions and ROM Locations

Table 2 shows the available HWAFFT routines with descriptions and respective addresses in ROM.

Table 2. Available HWAFFT Routines

Function Name Description VC5505 (PG1.4) C5505/C5515
ROM Address (PG2.0) ROM
Address
hwafft_br Int32 (32-bit) vector bit-reversal, Strict alignment requirement 0x00ff7342 0x00ff6cd6
hwafft_8pts 8-point FFT/IFFT, 1 double-stage, 1 single-stage 0x00ff7356 0x00ff6cea
hwafft_16pts 16-point FFT/IFFT, 2 double-stages, 0 single-stages 0x00ff7445 0x00ff6dd9
hwafft_32pts 32-point FFT/IFFT, 2 double-stages, 1 single-stage 0x00ff759b 0x00ff6f2f
hwafft_64pts 64-point FFT/IFFT, 3 double-stages, 0 single-stages 0x00ff78a4 0x00ff7238
hwafft_128pts 128-point FFT/IFFT, 3 double-stages, 1 single-stage 0x00ff7a39 0x00ff73cd
hwafft_256pts 256-point FFT/IFFT, 4 double-stages, 0 single-stages 0x00ff7c4a 0x00ff75de
hwafft_512pts 512-point FFT/IFFT, 4 double-stages, 1 single-stage 0x00ff7e48 0x00ff77dc
hwafft_1024pts 1024-point FFT/IFFT, 5 double-stages, 0 single-stages 0x00ff80c2 0x00ff7a56

SPRABB6B– June 2010 – Revised January 2013 FFT Implementation on the TMS320VC5505, TMS320C5505, and 15
Submit Documentation Feedback TMS320C5515 DSPs
Copyright © 2010–2013, Texas Instruments Incorporated
HWAFFT Software Interface www.ti.com

4.5 Project Configuration for Calling Functions from ROM

NOTE: To execute the HWAFFT routines from the ROM of the DSP, the programmer must satisfy
memory allocation restrictions for the data and scratch buffers. See the device-specific errata
for an explanation of the restrictions and workarounds:
• TMS320VC5505/VC5504 Fixed-Point DSP Silicon Errata (Silicon Revision 1.4)
[literature number SPRZ281]
• TMS320C5505/C5504 Fixed-Point DSP Silicon Errata (Silicon Revision 2.0)
[literature number SPRZ310]
• TMS320C5515/C5514 Fixed-Point DSP Silicon Errata (Silicon Revision 2.0)
[literature number SPRZ308]

The HWAFFT functions occupy approximately 4 KBytes of memory, so to conserve RAM they have been
placed in the DSP’s 128 KBytes of on-chip ROM. These functions are identical to and have the same
names as the functions stored in hwafft.asm, but they do not consume any RAM. In order to utilize these
HWAFFT routines in ROM, add the following lines to the bottom of the project’s linker CMD file and
remove the hwafft.asm file from the project (or exclude it from the build). When the project is rebuilt, the
HWAFFT functions will reference the ROM locations. The HWAFFT ROM locations are different between
VC5505 (PG1.4) and C5505/C5515 (PG2.0). ROM locations for both device families are shown in Table 2.
/*** Add the following code to the linker command file to call HWAFFT Routines from ROM ***/

/* HWAFFT Routines ROM Addresses */

/* (PG1.4) */
/*
_hwafft_br = 0x00ff7342;
_hwafft_8pts = 0x00ff7356;
_hwafft_16pts = 0x00ff7445;
_hwafft_32pts = 0x00ff759b;
_hwafft_64pts = 0x00ff78a4;
_hwafft_128pts = 0x00ff7a39;
_hwafft_256pts = 0x00ff7c4a;
_hwafft_512pts = 0x00ff7e48;
_hwafft_1024pts = 0x00ff80c2;
*/

/* HWAFFT Routines ROM Addresses */

/* (PG 2.0) */
_hwafft_br = 0x00ff6cd6;
_hwafft_8pts = 0x00ff6cea;
_hwafft_16pts = 0x00ff6dd9;
_hwafft_32pts = 0x00ff6f2f;
_hwafft_64pts = 0x00ff7238;
_hwafft_128pts = 0x00ff73cd;
_hwafft_256pts = 0x00ff75de;
_hwafft_512pts = 0x00ff77dc;
_hwafft_1024pts = 0x00ff7a56;

16 FFT Implementation on the TMS320VC5505, TMS320C5505, and SPRABB6B– June 2010 – Revised January 2013
TMS320C5515 DSPs Submit Documentation Feedback
Copyright © 2010–2013, Texas Instruments Incorporated
www.ti.com Simple Example to Illustrate the Use of the FFT Accelerator

5 Simple Example to Illustrate the Use of the FFT Accelerator

NOTE: To execute the HWAFFT routines from the ROM of the DSP, the programmer must satisfy
memory allocation restrictions for the data and scratch buffers. See the device-specific errata
for an explanation of the restrictions and workarounds:
• TMS320VC5505/VC5504 Fixed-Point DSP Silicon Errata (Silicon Revision 1.4)
[literature number SPRZ281]
• TMS320C5505/C5504 Fixed-Point DSP Silicon Errata (Silicon Revision 2.0)
[literature number SPRZ310]
• TMS320C5515/C5514 Fixed-Point DSP Silicon Errata (Silicon Revision 2.0)
[literature number SPRZ308]

The source code below demonstrates typical use of the HWAFFT for the 1024-point FFT and IFFT cases.
The HWAFFT Functions make use of Boolean flag variables to select between FFT and IFFT, Scale and
No Scale mode, and Data and Scratch output locations.
#define FFT_FLAG ( 0 ) /* HWAFFT to perform FFT */
#define IFFT_FLAG ( 1 ) /* HWAFFT to perform IFFT */
#define SCALE_FLAG ( 0 ) /* HWAFFT to scale butterfly output */
#define NOSCALE_FLAG ( 1 ) /* HWAFFT not to scale butterfly output */
#define OUT_SEL_DATA ( 0 ) /* Indicates HWAFFT output located in input data vector */
#define OUT_SEL_SCRATCH ( 1 ) /* Indicates HWAFFT output located in scratch vector */
Int32 *data;
Int32 *data_br;
Uint16 fft_flag;
Uint16 scale_flag;
Int32 *scratch;
Uint16 out_sel;
Int32 *result;

5.1 1024-Point FFT, Scaling Disabled

Compute 1024-point FFT with Scaling enabled: a ½ scale factor after every stage:
fft_flag = FFT_FLAG;
scale_flag = SCALE_FLAG;

data = <1024-point Complex input>;

/* Bit-Reverse 1024-point data, Store into data_br, data_br aligned to

12-least significant binary zeros*/
hwafft_br(data, data_br, DATA_LEN_1024); /* bit-reverse input data,
Destination buffer aligned */
data = data_br;

/* Compute 1024-point FFT, scaling enabled. */

out_sel = hwafft_1024pts(data, scratch, fft_flag, scale_flag);

if (out_sel == OUT_SEL_DATA) {
result = data;
}else {
result = scratch;
}

SPRABB6B– June 2010 – Revised January 2013 FFT Implementation on the TMS320VC5505, TMS320C5505, and 17
Submit Documentation Feedback TMS320C5515 DSPs
Copyright © 2010–2013, Texas Instruments Incorporated
Simple Example to Illustrate the Use of the FFT Accelerator www.ti.com

5.2 1024-Point IFFT, Scaling Disabled

Compute 1024-point IFFT with Scaling disabled:
fft_flag = IFFT_FLAG;
scale_flag = NOSCALE_FLAG;

data = <1024-point Complex input>;

/* Bit-Reverse 1024-point data, Store into data_br, data_br aligned to

12-least significant binary zeros */
hwafft_br(data, data_br, DATA_LEN_1024);
data = data_br;

/* Compute 1024-point IFFT, scaling disabled */

out_sel = hwafft_1024pts(data, scratch, fft_flag, scale_flag);

if (out_sel == OUT_SEL_DATA) {
result = data;
} else {
result = scratch;
}

18 FFT Implementation on the TMS320VC5505, TMS320C5505, and SPRABB6B– June 2010 – Revised January 2013
TMS320C5515 DSPs Submit Documentation Feedback
Copyright © 2010–2013, Texas Instruments Incorporated
www.ti.com Simple Example to Illustrate the Use of the FFT Accelerator

5.3 Graphing FFT Results in CCS4

Code Composer includes a graphing utility that makes visualization of the FFT operation quick and easy.
The Graph Utility is located in the CCSv4 window, under Tools → Graph → Single Time.
If the FFT Result is stored in scratch (OutSel = 1) and scratch is located at address 0x3000…
Plot the real part:

Figure 6. Graphing the Real Part of the FFT Result in CCS4

Plot the imaginary part:

Figure 7. Graphing the Imaginary Part of the FFT Result in CCS4

SPRABB6B– June 2010 – Revised January 2013 FFT Implementation on the TMS320VC5505, TMS320C5505, and 19
Submit Documentation Feedback TMS320C5515 DSPs
Copyright © 2010–2013, Texas Instruments Incorporated
FFT Benchmarks www.ti.com

6 FFT Benchmarks
Table 3 compares the FFT performance of the HWAFFT versus FFT computation using the CPU under
the following conditions:
• Core voltage = 1.05 V
• PLL = 60 MHz
• Power measurement condition:
– At room temperature only
– All peripherals are clock gated
– Measured at VDDC

Table 3. FFT Performance on HWAFFT vs CPU (Vcore = 1.05 V, PLL = 60 MHz)

FFT with HWA CPU (Scale) HWA versus CPU
Complex FFT FFT + BR (1) Cycles Energy/FFT FFT + BR (1) Cycles Energy/FFT x Times Faster x Times Energy
(nJ/FFT) (nJ/FFT) (Scale) Efficient (Scale)
8 pt 92 + 38 = 130 23.6 196 + 95 = 291 95.1 2.2 4
16 pt 115 + 55 = 170 32.1 344 + 117 = 461 157.1 2.7 4.9
32 pt 234 + 87 = 321 69.5 609 + 139 = 748 269.9 2.3 3.9
64 pt 285 + 151 = 436 98.5 1194 + 211 = 1405 531.7 3.2 5.4
128 pt 633 + 279 = 912 219.2 2499 + 299 = 2798 1090.4 3.1 5
256 pt 1133 + 535 = 1668 407.2 5404 + 543 = 5947 2354.2 3.6 5.8
512 pt 2693 + 1047 = 3740 939.7 11829 + 907 = 12736 5097.5 3.4 5.4
1024 pt 5244 + 2071 = 7315 1836.2 25934 + 1783 = 27717 11097.9 3.8 6
(1)
BR = Bit Reverse

In summary, Table 3 shows that for the test conditions used, HWAFFT is 4 to 6 times more energy
efficient and 2.2 to 3.8 times faster than the CPU. Table 4 compares FFT performance of the accelerator
versus FFT computation using the CPU under the following conditions:
• Core voltage = 1.3 V
• PLL = 100 MHz
• Power measurement condition:
– At room temperature only
– All peripherals are clock gated
– Measured at VDDC

Table 4. FFT Performance on HWAFFT vs CPU (Vcore = 1.3 V, PLL = 100 MHz)
FFT with HWA CPU (Scale) HWA versus. CPU
Complex FFT FFT + BR (1) Cycles Energy/FFT FFT + BR (1) Cycles Energy/FFT x Times Faster x Times Energy
(nJ/FFT) (nJ/FFT) (Scale) Efficient (Scale)
8 pt 92 + 38 = 130 36.3 196 + 95 = 291 145.9 2.2 4
16 pt 115 + 55 = 170 49.3 344 + 117 = 461 241 2.7 4.9
32 pt 234 + 87 = 321 106.9 609 + 139 = 748 414 2.3 3.9
64 pt 285 + 151 = 436 151.3 1194 + 211 = 1405 815.7 3.2 5.4
128 pt 633 + 279 = 912 336.8 2499 + 299 = 2798 1672.9 3.1 5
256 pt 1133 + 535 = 1668 625.6 5404 + 543 = 5947 3612.9 3.6 5.8
512 pt 2693 + 1047 = 3740 1442.8 11829 + 907 = 12736 7823.8 3.4 5.4
1024 pt 5244 + 2071 = 7315 2820.6 25934 + 1783 = 27717 17032.4 3.8 6
(1)
BR = Bit Reverse

In summary, Table 4 shows that for the test conditions used, HWAFFT is 4 to 6 times more energy
efficient and 2.2 to 3.8 times faster than the CPU.

20 FFT Implementation on the TMS320VC5505, TMS320C5505, and SPRABB6B– June 2010 – Revised January 2013
TMS320C5515 DSPs Submit Documentation Feedback
Copyright © 2010–2013, Texas Instruments Incorporated
www.ti.com Description of Open Source FFT Example Software

7 Description of Open Source FFT Example Software

An example application of the HWAFFT used in a real-time audio filter is available on the Open Source
C5505 eZdsp website (http://www.code.google.com/p/c5505-ezdsp). The zip file named “VC5505 FFT
Filter Demo” contains a Code Composer 4 Project and source code that implements a real-time low-pass
filter on the VC5505 eZdsp USB Stick. Low-pass filtering is achieved through multiplication with a filter in
the Frequency Domain. Recall that multiplication in the frequency domain is equivalent to convolution in
the time domain.
In this demo, 16-bit stereo samples are captured by the AIC3204 codec at a sampling frequency of 48 kHz
and copied to the DSP memory with the DMA over the Inter-IC Sound (I2S) bus. Samples from the left and
right channel are collected in separate ping-pong buffers. When the buffer becomes full a DMA interrupt
updates the ping-pong buffer and triggers the FFT filter to convolve the new block of samples with the low-
pass filter.
Filtering is performed in three steps:
1. Use the HWAFFT to calculate the FFT of the input block of samples from the ping-pong buffers.
2. Multiply this complex FFT result with the pre-computed FFT result of the filter coefficients.
Note: The FFT result of the filter coefficients is computed once during program initialization and stored
for reuse.
3. Calculate the IFFT of that product on the HWAFFT.
Because a stream of samples is constantly arriving at the codec and block processing is utilized to filter
the signal, the Constant-Overlap-and-Add (COLA) method is implemented to output a continuous, glitch-
free signal to the codec. Finally, the resulting block of filtered and overlapped samples is transferred back
to the codec for output with the DMA over the I2S bus.
This demo provides you control over FFT Lengths (from 8 to 1024 points) and Filter Lengths (from 7 taps
to 511 taps) for a thorough comparison. Additionally, simulation modes are available for using ideal
sinusoidal signals (stored in memory) as inputs to the FFT Filter Demo.
The block diagram in Figure 8 shows the data flow for one channel of the FFT Filter Demo. When
processing stereo input (separate left and right channels), this data flow is duplicated for each channel.

Filter
Coeffs

FFT

InBuf 1 OutBuf 1
Overlap
CPLX
FFT IFFT &
From Codec or MUL Add To Codec
Waveforms in InBuf 2 OutBuf 2
Memory

Figure 8. FFT Filter Demo Block Diagram

SPRABB6B– June 2010 – Revised January 2013 FFT Implementation on the TMS320VC5505, TMS320C5505, and 21
Submit Documentation Feedback TMS320C5515 DSPs
Copyright © 2010–2013, Texas Instruments Incorporated
Computation of Large (Greater Than 1024-Point) FFTs www.ti.com

8 Computation of Large (Greater Than 1024-Point) FFTs

The HWAFFT can perform up to 1024-point complex FFTs/IFFTs at a maximum, but if larger FFT sizes
(i.e. 2048-point) are required, the DSP core can be used to compute extra Radix-2 stages that are too
large for the HWAFFT to handle.
Recall the Radix-2 DIT equations:
1
X (k ) = (X (k ) + W k X (k )), k = 0 to N / 2 - 1
2 even N odd (11)
and
1
X (k + N / 2) = (X (k ) - W k X (k )), k = 0 to N / 2 - 1
2 even N odd (12)

8.1 Procedure for Computing Large FFTs

The procedure for computing an additional Radix-2 DIT stage on the CPU is outlined:
• Split the input signal into even and odd indexed signals, Xeven and Xodd.
• Call N/2 point FFTs for the even and odd indexed inputs.
• Complex Multiply the Xodd FFT results with the decimated twiddle factors for that stage.
• Add the Xodd * Twiddle product to Xeven to find the first half of the FFT result.
• Subtract Xodd * Twiddle to find the second half of the FFT result.

8.2 Twiddle Factor Computation

The HWAFFT stores 512 complex twiddle factors enabling FFT/IFFT computations up to 1024 points.
Recall Equation 5 states that only twiddle factors from 0 to N/2 are needed. To compute FFT/IFFTs larger
than 1024 points, you must supply N/2 complex twiddle factors, where N is the FFT length (powers of 2).
The following MATLAB code creates real and imaginary parts of the twiddle factors for any N:
N = 2048;
n = 0:(N/2-1);
twid_r = cos(2*pi*n/N);
twid_i = -sin(2*pi*n/N);

8.3 Bit-Reverse Separates Even and Odd Indexes

A nice property of the bit-reversal process is the automatic separation of odd-indexed data from even-
indexed data. Before the bit-reverse, even indexes have a 0 in the least significant bit and odd indexes
have a 1 in the least significant bit. After the bit-reverse, even indexes have a 0 in the most significant bit,
and odd indexes have a 1 in the most significant bit. Therefore, all even indexed data resides in the first
half of the bit-reversed vector, and all odd indexed data resides in the second half of the bit-reversed
vector. This process meets two needs: separation of even and odd indexed-data vectors and bit-reversing
both vectors.

8.4 2048-point FFT Source Code

The following C source code demonstrates a 2048-point FFT using this approach. Two 1024-point FFTs
are computed on the HWAFFT, and a final Radix-2 stage is performed on the CPU to generate a 2048-
point FFT result:
#define FFT_FLAG ( 0 ) /* HWAFFT to perform FFT */
#define IFFT_FLAG ( 1 ) /* HWAFFT to perform IFFT */
#define SCALE_FLAG ( 0 ) /* HWAFFT to scale butterfly output */
#define NOSCALE_FLAG ( 1 ) /* HWAFFT not to scale butterfly output */
#define OUT_SEL_DATA ( 0 ) /* Indicates HWAFFT output located in input data vector */
#define OUT_SEL_SCRATCH ( 1 ) /* Indicates HWAFFT output located in scratch vector */
#define DATA_LEN_2048 ( 2048 )
#define TEST_DATA_LEN (DATA_LEN_2048)

// Static Memory Allocations and Alignment:

#pragma DATA_SECTION(data_br_buf, "data_br_buf");

22 FFT Implementation on the TMS320VC5505, TMS320C5505, and SPRABB6B– June 2010 – Revised January 2013
TMS320C5515 DSPs Submit Documentation Feedback
Copyright © 2010–2013, Texas Instruments Incorporated
www.ti.com Computation of Large (Greater Than 1024-Point) FFTs
#pragma DATA_ALIGN (data_br_buf, 4096);
// Align 2048-pt bit-reverse dest vector to byte addr w/ 13 least sig zeros
Int32 data_br_buf[TEST_DATA_LEN];

#pragma DATA_SECTION(data_even_buf, "data_even_buf");

Int32 data_even_buf[TEST_DATA_LEN/2];

#pragma DATA_SECTION(data_odd_buf, "data_odd_buf");

Int32 data_odd_buf[TEST_DATA_LEN/2];

#pragma DATA_SECTION(scratch_even_buf, "scratch_even_buf");

Int32 scratch_even_buf[TEST_DATA_LEN/2];

#pragma DATA_SECTION(scratch_odd_buf, "scratch_odd_buf");

Int32 scratch_odd_buf[TEST_DATA_LEN/2];

// Function Prototypes:
Int32 CPLX_Mul(Int32 op1, Int32 op2);
// Yr = op1_r*op2*r - op1_i*op2_i, Yi = op1_r*op2_i + op1_i*op2_r
Int32 CPLX_Add(Int32 op1, Int32 op2, Uint16 scale_flag);
// Yr = 1/2 * (op1_r + op2_r), Yi = 1/2 *(op1_i + op2_i)
Int32 CPLX_Subtract(Int32 op1, Int32 op2, Uint16 scale_flag);
// Yr = 1/2 * (op1_r - op2_r), Yi = 1/2 *(op1_i - op2_i)

// Declare Variables
Int32 *data_br;
Int32 *data;
Int32 *data_even, *data_odd;
Int32 *scratch_even, *scratch_odd;
Int32 *twiddle;
Int32 twiddle_times_data_odd;
Uint16 fft_flag;
Uint16 scale_flag;
Uint16 out_sel;
Uint16 k;

// Assign pointers to static memory allocations

data_br = data_br_buf;
data_even = data_even_buf;
data_odd = data_odd_buf;
scratch_even = scratch_even_buf;
scratch_odd = scratch_odd_buf;
twiddle = twiddle_buf; // 1024-pt Complex Twiddle Table
data = invec_fft_2048pts; // 2048-pt Complex Input Vector

// HWAFFT flags:
fft_flag = FFT_FLAG; // HWAFFT to perform FFT (not IFFT)
scale_flag = SCALE_FLAG; // HWAFFT to scale by 2 after each butterfly stage

// Bit-reverse input data for DIT FFT calculation

hwafft_br(data, data_br, DATA_LEN_2048);
// data_br aligned to log2(4*2048) = 13 zeros in least sig bits
data = data_br;

// Split data into even-indexed data & odd-indexed data

// data is already bit-reversed, so even-indexed data = first half & odd-
indexed data = second half
for(k=0; k<DATA_LEN_2048/2; k++)
{
data_even[k] = data[k];
data_odd[k] = data[k+DATA_LEN_2048/2];
}

// 1024-pt FFT the even data on the FFT Hardware Accelerator

out_sel = hwafft_1024pts(data_even, scratch_even, fft_flag, scale_flag);
if(out_sel == OUT_SEL_SCRATCH) data_even = scratch_even;

SPRABB6B– June 2010 – Revised January 2013 FFT Implementation on the TMS320VC5505, TMS320C5505, and 23
Submit Documentation Feedback TMS320C5515 DSPs
Copyright © 2010–2013, Texas Instruments Incorporated
Computation of Large (Greater Than 1024-Point) FFTs www.ti.com

// 1024-pt FFT the odd data on the FFT Hardware Accelerator

out_sel = hwafft_1024pts(data_odd, scratch_odd, fft_flag, scale_flag);
if(out_sel == OUT_SEL_SCRATCH) data_odd = scratch_odd;

// Combine the even and odd FFT results with a final Radix-2 Butterfly stage on the CPU
for(k=0; k<DATA_LEN_2048/2; k++) // Computes 2048-point FFT
{
// X(k) = 1/2*(X_even[k] + Twiddle[k]*X_odd(k))
// X(k+N/2) = 1/2*(X_even[k] - Twiddle[k]*X_odd(k))
// Twiddle[k]*X_odd(k):

twiddle_times_data_odd = CPLX_Mul(twiddle[k], data_odd[k]);

// X(k):
data[k] = CPLX_Add(data_even[k], twiddle_times_data_odd, SCALE_FLAG); // Add then scale by 2

// X(k+N/2):
data[k+DATA_LEN_2048/2] = CPLX_Subtract(data_even[k], twiddle_times_data_odd, SCALE_FLAG);
//Sub then scale
}

result = data; //2048-pt FFT result

/* END OF 2048-POINT FFT SOURCE CODE */

24 FFT Implementation on the TMS320VC5505, TMS320C5505, and SPRABB6B– June 2010 – Revised January 2013
TMS320C5515 DSPs Submit Documentation Feedback
Copyright © 2010–2013, Texas Instruments Incorporated
www.ti.com Appendix A Methods for Aligning the Bit-Reverse Destination Vector

9 Appendix A Methods for Aligning the Bit-Reverse Destination Vector

The optimized bit-reverse function hwafft_br requires the destination vector to be data aligned such that
the starting address of the destination vector, data_br, contains log2(4 * N) zeros in the least significant
bits of the binary address. There are a few different ways to force the linker map the bit-reverse
destination vector to an address with log2(4 * N) zeros in the least significant bits. Three different methods
are shown here. For further details, refer to the TMS320C55x C/C++ Compiler User’s Guide (SPRU280).

9.1 Statically Allocate Buffer at Beginning of Suitable RAM Block

NOTE: To execute the HWAFFT routines from the ROM of the DSP, the programmer must satisfy
memory allocation restrictions for the data and scratch buffers. See the device-specific errata
for an explanation of the restrictions and workarounds:
• TMS320VC5505/VC5504 Fixed-Point DSP Silicon Errata (Silicon Revision 1.4)
[literature number SPRZ281]
• TMS320C5505/C5504 Fixed-Point DSP Silicon Errata (Silicon Revision 2.0)
[literature number SPRZ310]
• TMS320C5515/C5514 Fixed-Point DSP Silicon Errata (Silicon Revision 2.0)
[literature number SPRZ308]

Place the buffer at the beginning of a DARAM or SARAM block with log2(4 * N) zeros in the least
significant bits of its byte address. For example, memory section DARAM2_3 below starts at address
0x0004000, which contains 14 zeros in the least significant bits of its binary address (0x0004000 =
0b0100 0000 0000 0000). Therefore, this address is a suitable bit-reverse destination vector for FFT
Lengths up to 4096-points because log2(4 * 4096) = 14.
In the Linker CMD File...
MEMORY
{
MMR (RWIX): origin = 0000000h, length = 0000c0h /* MMRs */
DARAM0 (RWIX): origin = 00000c0h, length = 001f40h /* on-chip DARAM 0, 4000 words */
DARAM1 (RWIX): origin = 0002000h, length = 002000h /* on-chip DARAM 1, 4096 words */
DARAM2_3 (RWIX): origin = 0004000h, length = 004000h /* on-chip DARAM 2_3, 8192 words */
DARAM4 (RWIX): origin = 0008000h, length = 002000h /* on-chip DARAM 4, 4096 words */
... (leaving out rest of memory sections)
}

SECTIONS
{
data_br_buf : > DARAM2_3 /* ADDR = 0x004000, Aligned to addr with 14 least-sig zeros */
}

SPRABB6B– June 2010 – Revised January 2013 FFT Implementation on the TMS320VC5505, TMS320C5505, and 25
Submit Documentation Feedback TMS320C5515 DSPs
Copyright © 2010–2013, Texas Instruments Incorporated
Appendix A Methods for Aligning the Bit-Reverse Destination Vector www.ti.com

9.2 Use the ALIGN Descriptor to Force log2(4 * N) Zeros in the Least Significant Bits
The ALIGN descriptor forces the alignment of a specific memory section, while providing the linker with
added flexibility to allocate sections across the entire DARAM or SARAM because no blocks are statically
allocated. It aligns the memory section to an address with log2(ALIGN Value) zeros in the least significant
bits of the binary address.
For example, the following code aligns data_br_buf to an address with 12 zeros in the least significant
bits, suitable for a 1024-point bit-reverse destination vector.
In the Linker CMD File...
MEMORY
{
MMR (RWIX): origin = 0000000h, length = 0000c0h /* MMRs */
DARAM (RWIX): origin = 00000c0h, length = 00ff40h /* on-chip DARAM 32 Kwords */
SARAM (RWIX): origin = 0010000h, length = 040000h /* on-chip SARAM 128 Kwords */
}

SECTIONS
{
data_br_buf : > DARAM ALIGN = 4096
/* 2^12 = 4096 , Aligned to addr with 12 least-sig zeros */
}

9.3 Use the DATA_ALIGN Pragma

The DATA_ALIGN pragma is placed in the source code where the vector is defined. The syntax is shown
below.
#pragma DATA_ALIGN (symbol, constant);
The DATA_ALIGN pragma aligns the symbol to an alignment boundary. The boundary is the value of the
constant in words. For example, a constant of 4 specifies a 64-bit alignment. The constant must be a
power of 2.
In this example, a constant of 2048 aligns the data_br_buf symbol to an address with 12 zeros in the least
significant bits, suitable for a 1024-point bit-reverse destination vector.
In the source file where data_br is declared (e.g. main.c)...
#pragma DATA_SECTION(data_br_buf, "data_br_buf");
#pragma DATA_ALIGN (data_br_buf, 2048);
Int32 data_br_buf[TEST_DATA_LEN];

26 FFT Implementation on the TMS320VC5505, TMS320C5505, and SPRABB6B– June 2010 – Revised January 2013
TMS320C5515 DSPs Submit Documentation Feedback
Copyright © 2010–2013, Texas Instruments Incorporated
www.ti.com

Appendix A Revision History

This revision history highlights the changes made to this document to make it a SPRABB6B revision.

Table 5. Revision History

See Revision
Added notes to satisfy memory allocation restrictions for the data and scratch FFT
Entire document
buffers before executing HWAFFT routines from the ROM of the DSP.

SPRABB6B– June 2010 – Revised January 2013 FFT Implementation on the TMS320VC5505, TMS320C5505, and 27
Submit Documentation Feedback TMS320C5515 DSPs
Copyright © 2010–2013, Texas Instruments Incorporated
 IMPORTANT NOTICE AND DISCLAIMER

TI PROVIDES TECHNICAL AND RELIABILITY DATA (INCLUDING DATASHEETS), DESIGN RESOURCES (INCLUDING REFERENCE
DESIGNS), APPLICATION OR OTHER DESIGN ADVICE, WEB TOOLS, SAFETY INFORMATION, AND OTHER RESOURCES “AS IS”
AND WITH ALL FAULTS, AND DISCLAIMS ALL WARRANTIES, EXPRESS AND IMPLIED, INCLUDING WITHOUT LIMITATION ANY
IMPLIED WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE OR NON-INFRINGEMENT OF THIRD
PARTY INTELLECTUAL PROPERTY RIGHTS.
These resources are intended for skilled developers designing with TI products. You are solely responsible for (1) selecting the appropriate
TI products for your application, (2) designing, validating and testing your application, and (3) ensuring your application meets applicable
standards, and any other safety, security, or other requirements. These resources are subject to change without notice. TI grants you
permission to use these resources only for development of an application that uses the TI products described in the resource. Other
reproduction and display of these resources is prohibited. No license is granted to any other TI intellectual property right or to any third
party intellectual property right. TI disclaims responsibility for, and you will fully indemnify TI and its representatives against, any claims,
damages, costs, losses, and liabilities arising out of your use of these resources.
TI’s products are provided subject to TI’s Terms of Sale (www.ti.com/legal/termsofsale.html) or other applicable terms available either on
ti.com or provided in conjunction with such TI products. TI’s provision of these resources does not expand or otherwise alter TI’s applicable
warranties or warranty disclaimers for TI products.

Mailing Address: Texas Instruments, Post Office Box 655303, Dallas, Texas 75265
Copyright © 2019, Texas Instruments Incorporated