AVR32718: AT32UC3 Series Software

AVR32718: AT32UC3 Series Software Framework
DSPLib
32-bit
1. Introduction
This application note describes the DSP Library from the AVR32® Software Framework. It details the main functions (prototype, algorithm and benchmark) of the DSP
library: FFT, convolution, FIR and partial IIR using GCC compiler.
All the source code (C code and assembly), software example and GCC and IAR
projects are released in the AVR32 UC3 Software Framework.
1.1
Microcontrollers
Application Note
References
• AVR32 UC3 Software Framework: This framework provides software drivers,
libraries and application examples to build any application for AVR32 UC3.
devices.http://www.atmel.com/dyn/products/tools_card.asp?tool_id=4192
• AVR32 Architecture Manual:
http://www.atmel.com/dyn/resources/prod_documents/doc32000.pdf
32076A–AVR32–11/07
2. Radix-4 decimate in time complex FFT
2.1
Description
This function computes a complex FFT from an input signal. It uses the Radix-4 “Decimate In
Time” algorithm and does not perform a calculation “in place” which means that the input buffer
has to be different from the output buffer.Function prototype
void dspXX_trans_complexfft(
dspXX_complex_t *vect1,
dspXX_t *vect2,
int nlog);
where XX corresponds to the number of bits of a basic data element (i.e. 16 or 32).
2.1.1
Arguments
This function takes three parameters: the output buffer, the input buffer and a value corresponding to the size of those buffers.
The output buffer (vect1) is a pointer on a complex vector of 2^nlog elements.
The input buffer (vect2) is a pointer on a real vector of 2^nlog elements.
The size argument (nlog) is in fact the base-2-logarithm of the size of the input vector. (nlog fits
in [2, 4, 6, …, 28])
2.1.2
Algorithm
Following is the algorithm used to implement the radix-4 DIT complex FFT. The optimized version is based on this algorithm but can differ in certain points due to the instruction set of the
target:
size = 1 << nlog
FOR r FROM 0 TO size-1 STEP 4 DO
Butterfly_zero_only_real_and_bit_reversing(vect1, vect2, r)
END
FOR stage FROM 1 TO nlog/2 DO
m = 4 ^ stage
FOR r FROM 0 TO size-1 STEP m DO
Butterfly_zero(vect1, r)
END
FOR j FROM 1 TO m / 4 - 1 DO
Comput_twiddle_factors(e, e2, e3, j / m)
FOR r FROM 0 TO size-1 STEP m DO
Butterfly(vect1, r, j, e, e2, e3)
END
END
END
2
AVR32718
32076A–AVR32–11/07
AVR32718
2.1.3
Notes
Interruptibility: the code is interruptible.
In-place computation is not allowed.
This function uses a static twiddle factors table raw-coded in the file “BASIC/TRANSFORMS/dspXX_twiddle_factors.h”. To generate those factors, you can use the script called
“tf_gen.sci” and execute it with Scilab.
To avoid overflowing values, the resulting vector amplitude is scaled by 2^nlog.
All the vectors have to be 32-bit aligned.
2.2
2.2.1
Benchmark
Benchmark routine
All these functions have been benchmarked on an avr32-uc3a0512 target. The programs have
been compiled with avr32-gcc (4.0.2-atmel.1.0.0) with the –O3 optimization option and have
been stored in FLASH memory. The fixed-point format used is the Q1.15 format for the 16-bit
data and the Q1.31 format for the 32-bit data.
The benchmark process has been performed with the same input signal for all those functions
and compared with a reference’s signal computed with a mathematic tool using floating point.
The input signal is a combination of one sine and one cosine. The sine oscillating at 400Hz and
the cosine at 2KHz. Those signals have been multiplied and sampled at 40KHz.
>
Complex
FFT
Input signal
2.2.2
Signal resulting and formatted (FFT)
Result
Here are tables of the main values of the benchmark results. All those values correspond to the
best performances of the functions and are obtained with different compilation options. For more
information, please refer to the complete benchmark result table in annexes.
3
32076A–AVR32–11/07
2.2.2.1
16-bit radix-4 D.I.T. complex FFT: generic
Concerned file path: /BASIC/TRANSFORMS/dsp16_complex_fft_generic.c
Lowest Error
Lowest cycle count
Fastest computation
at 60 MHz
Amplitude
average
Max. amplitude
Lowest Algorithm’s size
in memory
1.1 Kbytes
64-points
6,296
108.2us
1.58e-5
6.53e-5
256-points
33,723
578.0us
1.69e-5
8.80e-5
1.3 Kbytes
1024-points
169,006
2.90ms
1.67e-5
12.31e-4
2.0 Kbytes
4096-points
812,321
13.90ms
1.52e-5
14.60e-4
5.0 Kbytes
More details on Table 1.1.1 in annexes
2.2.2.2
16-bit radix-4 D.I.T. complex FFT: avr32-uc3 optimized
Concerned file path: /BASIC/TRANSFORMS/dsp16_complex_fft_avr32uc3.c
Lowest Error
Lowest cycle count
Fastest computation
at 60 MHz
Amplitude
average
Max. amplitude
Lowest Algorithm’s size
in memory
64-points
2,611
44.4 us
1.63e-5
6.53e-5
710 bytes
256-points
13,661
232.2 us
1.68e-5
7.46e-5
902 bytes
1024-points
67,671
1.15 ms
1.69e-5
1.02e-4
1.6 Kbytes
4096-points
322,897
5.49 ms
1.58e-5
1.18e-4
4.6 Kbytes
Warning: this function is only compatible with Q1.15 numbers.
Note: this function needs 72 bytes of memory for the stack.
More details on Table 1.1.2 in annexes
2.2.2.3
32-bit radix-4 D.I.T. complex FFT: generic
Concerned file path: /BASIC/TRANSFORMS/dsp32_complex_fft_generic.c
Lowest Error
Lowest cycle count
Fastest computation
at 60 MHz
Amplitude
average
Max. amplitude
Lowest Algorithm’s size
in memory
64-points
13,206
225.2us
6.0e-10
5.7e-9
2.0 Kbytes
256-points
74,297
1.27us
3.0e-10
4.8e-9
2.4 Kbytes
1024-points
383,212
6.53ms
3.0e-10
6.1e-9
3.9 Kbytes
More details on Table 1.2.1 in annexes
4
AVR32718
32076A–AVR32–11/07
AVR32718
3. Convolution
3.1
Description
This function performs a linear convolution between two discrete sequences.
3.1.1
Function prototype
void dspXX_vect_conv(
dspXX_t *vect1,
dspXX_t *vect2,
int vect2_size,
dspXX_t *vect3,
int vect3_size);
where XX corresponds to the number of bits of a basic data element (i.e. 16 or 32).
3.1.2
Arguments
This function takes five parameters: the output buffer, the two discrete sequences and their
respective sizes.
• The output buffer (vect1) is a pointer on a real vector of (vect2_size + vect3_size - 1)
elements.
• The first input buffer (vect2) is a pointer on a real vector of vect2_size elements.
• The first size argument (vect2_size) is the length of the first input buffer (vect2_size ∈ [8, 9,
10, …]).
• The second input buffer (vect3) is a pointer on a real vector of vect3_size elements.
• The second size argument (vect3_size) is the length of the second input buffer (vect3_size
fits in [8, 9, 10, …]).
3.1.3
Requirements
This function requires 3 modules:
Module name
Function name
Concerned file path
Zero Padding
dspXX_vect_zeropad
/BASIC/VECTORS/zero_padding.c
Copy
dspXX_vect_copy
/BASIC/VECTORS/copy.c
Partial Convolution
dspXX_vect_convpart
/BASIC/VECTORS/convolution_partial.c
The output buffer of the function has to have at least a length of N + 2*M – 2 elements because of intern computations, where N is the length of the largest input buffer and M,
the length of the smallest input buffer.
5
32076A–AVR32–11/07
3.1.4
Algorithm
Following is the algorithm used to implement the convolution product. The optimized version is
based on this algorithm but can differ in certain points due to the instruction set of the target:
IF vect2_size >= vect3_size THEN
vect1 =
0000…0000
vect2
0000…0000
vect3_size – 1
vect2_size
vect3_size – 1
Partial_convolution(vect1, vect1, vect2_size + 2*(vect3_size – 1), vect3, vect3_size)
ELSE
vect1 =
0000…0000
vect3
0000…0000
vect2_size – 1
vect3_size
vect2_size – 1
Partial_convolution(vect1, vect1, vect3_size + 2*(vect2_size – 1), vect2, vect2_size)
END
3.1.5
Notes
• Interruptibility: the code is interruptible.
• Due to its implementation, the dsp16-avr32-uc3 optimized version of the FIR requires a
length of 4*m elements for the largest input discrete sequence and the output buffer (vect1)
has to have a length of 4*n elements to avoid overflows.
• The input discrete sequences have to be scaled to avoid overflowing values.
• All the vectors have to be 32-bit aligned.
3.2
3.2.1
Benchmark
Benchmark routine
All these functions have been benchmarked on an avr32-uc3a0512 target. The programs have
been compiled with avr32-gcc (4.1.2-atmel.1.0.0) with the –O3 optimization option and have
been stored in FLASH memory. The fixed-point format used is the Q1.15 format for the 16-bit
data and the Q1.31 format for the 32-bit data.
The benchmark process has been performed with the same input signal and impulse response
for all those functions and compared with a reference’s signal computed with a mathematic tool
using floating point.
The first input signal is a sine oscillating at 433Hz and the second input signal is a cosine
oscillating at 2KHz. Those signals are sampled at 40KHz.
6
AVR32718
32076A–AVR32–11/07
AVR32718
>
1st input signal
FIR filter
Resulting signal
2nd input signal
3.2.2
Result
Here are tables of the main values of the benchmark results. All those values correspond to the
best performances of the functions and are obtained with different compilation options. For more
information, please refer to the complete benchmark result table in annexes.
Concerned file path: /BASIC/VECTORS/convolution.c
3.2.2.1
16-bit Convolution: generic
Algorithm’s size in memory: 2.2 Kbytes.
Length of the first input signal: 64 elements.
Lowest Error
Lowest cycle
count
Fastest computation at 60
MHz
Amplitude average
32-points
23,524
408.5us
2.0e-5
4.5e-5
64-points
57,757
1.00ms
1.8e-5
4.7e-5
128-points
86,752
1.50ms
1.8e-5
4.4e-5
256-points
144,736
2.51ms
1.5e-5
4.8e-5
Max. amplitude
More details on Table 2.1.1 in annexes
7
32076A–AVR32–11/07
3.2.2.2
16-bit Convolution: avr32-uc3 optimized
Algorithm’s size in memory: 950 bytes.
Length of the first input signal: 64 elements.
Lowest Error
Lowest cycle
count
Fastest computation at 60
MHz
Amplitude average
32-points
8,248
151.2us
2.0e-5
4.5e-5
64-points
19,087
349.1us
1.8e-5
4.7e-5
128-points
28,532
521.8us
1.8e-5
4.4e-5
256-points
47,412
866.9us
1.5e-5
4.8e-5
Max. amplitude
More details on Table 2.1.2 in annexes
3.2.2.3
32-bit Convolution: generic
Algorithm’s size in memory: 3.3 Kbytes.
Length of the first input signal: 64 elements.
Lowest Error
Lowest cycle
count
Fastest computation at 60
MHz
Amplitude average
32-points
42,572
729.8us
0.4e-9
2.1e-9
64-points
109,179
1.87ms
0.4e-9
1.7e-9
128-points
163,968
2.81ms
0.5e-9
1.6e-9
256-points
273,536
4.69ms
0.6e-9
2.7e-9
Max. amplitude
More details on Table 2.2.1 in annexes
3.2.2.4
32-bit Convolution: avr32-uc3 optimized
Algorithm’s size in memory: 1.5 Kbytes.
Length of the first input signal: 64 elements.
Lowest Error
Lowest cycle
count
Fastest computation at 60
MHz
Amplitude average
Max. amplitude
32-points
19,958
340.2us
0.5e-9
2.1e-9
64-points
50,501
860.4us
0.5e-9
1.7e-9
128-points
75,722
1.29ms
0.6e-9
2.4e-9
256-points
126,154
2.15ms
0.7e-9
2.7e-9
More details on Table 2.2.2 in annexes
8
AVR32718
32076A–AVR32–11/07
AVR32718
4. FIR Filter (alias Partial Convolution)
4.1
Description
This function computes a real FIR filter using the impulse response of the desire filter onto a
fixed-length signal.
4.1.1
Function prototype
void dspXX_filt_fir(
void dspXX_vect_convpart(
dspXX_t *vect1,
dspXX_t *vect1,
dspXX_t *vect2,
dspXX_t *vect2,
int size,
int vect2_size,
dspXX_t *h,
dspXX_t *vect3,
int h_size);
int vect3_size);
where XX corresponds to the number of bits of a basic data element (i.e. 16 or 32).
4.1.2
Arguments
This function takes five parameters: the output buffer, the input buffer, its size, the impulse
response of the filter and its size.
The output buffer (vect1) is a pointer on a real vector of (size - h_size + 1) elements.
The input buffer (vect2) is a pointer on a real vector of size elements.
The size argument (size) is the length of the input buffer (size fits in [4, 8, 12, …]).
The impulse response of the filter (h) is a pointer on a real vector of h_size elements.
The size argument (h_size) is the length of the impulse response of the filter (h_size fits in [8, 9,
10, …]).
4.1.3
Requirements
This function requires one module:
Module name
Function name
Concerned file path
Partial Convolution
dspXX_vect_convpart
/BASIC/VECTORS/convolution_partial.c
9
32076A–AVR32–11/07
4.1.4
Algorithm
Following is the algorithm used to implement the FIR filter. The optimized version is based on
this algorithm but can differ in certain points due to the instruction set of the target:
FOR j FROM 0 TO size - h_size + 1 DO
sum = 0
FOR i FROM 0 TO h_size DO
sum += vect2[i] * h[h_size - i - 1]
END
vect1[j] = sum >> DSPXX_QB
END
4.1.5
Notes
• Interruptibility: the code is interruptible.
• Due to its implementation, for the dsp16-avr32-uc3 optimized version of the FIR, the output
buffer (vect1) has to have a length of 4*n elements to avoid overflows.
• The impulse response of the filter has to be scaled to avoid overflowing values.
• All the vectors have to be 32-bit aligned.
4.2
4.2.1
Benchmark
Benchmark routine
All these functions have been benchmarked on an avr32-uc3a0512 target. The programs have
been compiled with avr32-gcc (4.1.2-atmel.1.0.0) with the –O3 optimization option and have
been stored in FLASH memory. The fixed-point format used is the Q1.15 format for the 16-bit
data and the Q1.31 format for the 32-bit data.
The benchmark process has been performed with the same input signal and impulse response
for all those functions and compared with a reference’s signal computed with a mathematic tool
using floating point.
The input signal is a combination of one sine and one cosine. The sine oscillating at 400Hz and
the cosine at 2KHz. Those signals have been multiplied and sampled at 40KHz.
The impulse response describes a low-pass filter with a cutoff frequency equal to 400Hz.
10
AVR32718
32076A–AVR32–11/07
AVR32718
>
Input signal
FIR filter
Resulting signal
Impulse response of the filter
4.2.2
Result
Here are tables of the main values of the benchmark results. All those values correspond to the
best performances of the functions and are obtained with different compilation options. For more
information, please refer to the complete benchmark result table in annexes.
4.2.2.1
16-bit FIR filter: generic
Concerned file path: /BASIC/VECTORS/dsp16_convpart_generic.c
Algorithm’s size in memory: 2.0 Kbytes.
Number of Taps: 24.
Lowest Error
Lowest cycle
count
Fastest computation at 60
MHz
Amplitude average
64-points
7,424
128.0us
2.27e-5
9.46e-5
256-points
41,793
720.0us
2.22e-5
9.46e-5
512-points
87,617
1.51ms
2.23e-5
9.46e-5
1024-points
179,265
3.09ms
2.21e-5
9.46e-5
Max. amplitude
More details on Table 3.1.1 in annexes
4.2.2.2
16-bit FIR filter: avr32-uc3 optimized
Concerned file path: /BASIC/VECTORS/dsp16_convpart_avr32uc3.c
Algorithm’s size in memory: 770 bytes.
Number of Taps: 24.
Lowest Error
Lowest cycle
count
Fastest computation at 60
MHz
Amplitude average
64-points
2,439
44.3us
2.27e-5
9.46e-5
256-points
12,712
230.7us
2.22e-5
9.46e-5
512-points
26,408
479.2us
2.23e-5
9.46e-5
1024-points
53,800
976.3us
2.21e-5
9.46e-5
Max. amplitude
More details on Table 3.1.2 in annexes
11
32076A–AVR32–11/07
4.2.2.3
32-bit FIR filter: generic
Concerned file path: /BASIC/VECTORS/dsp32_convpart_generic.c
Algorithm’s size in memory: 3.1 Kbytes.
Number of Taps: 24.
Lowest Error
Lowest cycle
count
Fastest computation at 60
MHz
Amplitude average
64-points
13,984
239.4us
2.1e-9
1.24e-8
256-points
79,073
1.35ms
2.3e-9
1.74e-8
512-points
165,857
2.84ms
2.6e-9
2.31e-8
1024-points
339,425
5.81ms
3.7e-9
2.84e-8
Max. amplitude
More details on Table 3.2.1 in annexes
4.2.2.4
32-bit FIR filter: avr32-uc3 optimized
Concerned file path: /BASIC/VECTORS/dsp32_convpart_avr32uc3.c
Algorithm’s size in memory: 1.3 Kbytes.
Number of Taps: 24.
Lowest Error
Lowest cycle
count
Fastest computation at 60
MHz
Amplitude average
Max. amplitude
64-points
6,479
110.2us
2.1e-9
1.24e-8
256-points
36,432
619.0us
2.3e-9
1.24e-8
512-points
76,368
1.30ms
2.6e-9
2.31e-8
1024-points
156,240
2.65ms
3.7e-9
2.84e-8
More details on Table 3.2.2 in annexes
12
AVR32718
32076A–AVR32–11/07
AVR32718
5. Partial IIR Filter
5.1
Description
This function computes a real IIR filter using the impulse response of the desire filter onto a
fixed-length signal.
5.1.1
Function prototype
void dspXX_filt_iir(
dspXX_t *vect1,
dspXX_t *vect2,
int size,
dspXX_t *num,
int num_size,
dspXX_t *den,
int den_size,
int num_prediv,
int den_prediv);
where XX corresponds to the number of bits of a basic data element (i.e. 16 or 32).
5.1.2
Arguments
This function takes five parameters: the output buffer, the input buffer, its size, the coefficients of
the filter, theirs sizes and a coefficient’s predivisor.
The output buffer (vect1) is a pointer on a real vector of (size - num_size + 1) elements.
The input buffer (vect2) is a pointer on a real vector of size elements.
The size argument (size) is the length of the input buffer (size fits in [4, 5, 6, 7, …]).
The numerator’s coefficients argument of the filter (num) is a pointer on a real vector of
num_size elements.
The size argument (num_size) is the length of the numerator’s coefficients of the filter (num_size
fits in [1, 2, 3, …]).
The denominator’s coefficients argument of the filter (den) is a pointer on a real vector of
den_size elements.
The size argument (den_size) is the length of the denominator’s coefficients of the filter
(den_size fits in [1, 2, 3, …]).
The predivisors (num_prediv and den_prediv) are used to scale down the denominator/numerator’s coefficients of the filter in order to avoid overflow values. So when you use this feature, you
have to prescale manually the denominator/numerator’s coefficients by 2^prediv else leave this
field to 0.
13
32076A–AVR32–11/07
5.1.3
Algorithm
Following is the algorithm used to implement the IIR filter. The optimized version is based on this
algorithm but can differ in certain points due to the instruction set of the target:
// Initialization of the vect1 coefficients
FOR n FROM 0 TO den_size - 1 DO
sum1 = 0
FOR m FROM 0 TO num_size - 1 DO
sum1 += num[m] * vect2[n + num_size - m - 1]
END
sum2 = 0
FOR m FROM 1 TO n DO
sum2 += den[m] * vect1[n - m]
END
vect1[n] = (sum1 – (sum2 << prediv)) >> DSPXX_QB
END
FOR n FROM n TO size – num_size DO
sum1 = 0
FOR m FROM 0 TO num_size - 1 DO
sum1 += num[m] * vect2[n + num_size - m - 1]
END
sum2 = 0
FOR m FROM 1 TO den_size - 1 DO
sum2 += den[m] * vect1[n - m]
END
vect1[n] = (sum1 – (sum2 << prediv)) >> DSPXX_QB
END
5.1.4
Notes
• Interruptibility: the code is interruptible.
• Due to its implementation, for the dsp16-avr32-uc3 optimized version of the FIR, the output
buffer (vect1) has to have a length of 4*n elements to avoid overflows.
• The impulse response of the filter has to be scaled to avoid overflowing values.
• All the vectors have to be 32-bit aligned.
• The first denominator’s coefficient have to be equal to 1 / (2^prediv).
• The predivisor (prediv) must be lower or equals to the constant DSPXX_QB.
14
AVR32718
32076A–AVR32–11/07
AVR32718
5.2
5.2.1
Benchmark
Benchmark routine
All these functions have been benchmarked on an avr32-uc3a0512 target. The programs have
been compiled with avr32-gcc (4.1.2-atmel.1.0.0) with the –O3 optimization option and have
been stored in FLASH memory. The fixed-point format used is the Q1.15 format for the 16-bit
data and the Q1.31 format for the 32-bit data.
The benchmark process has been performed with the same input signal and impulse response
for all those functions and compared with a reference’s signal computed with a mathematic tool
using floating point.
The input signal is a combination of one sine and one cosine. The sine oscillating at 400Hz and
the cosine at 4KHz. Those signals have been added together and sampled at 8KHz.
The filter used is a low-pass Butterworth filter with a cutoff frequency equal to 2KHz.
15
32076A–AVR32–11/07
5.2.2
Result
Here are tables of the main values of the benchmark results. All those values correspond to the
best performances of the functions and are obtained with different compilation options. For more
information, please refer to the complete benchmark result table in annexes.
5.2.2.1
16-bit IIR filter: generic
Concerned file path: /BASIC/FILTERING/dsp16_iir_generic.c
Algorithm’s size in memory: 266 bytes.
Order of the filter: 7
Lowest Error
Lowest cycle
count
Fastest computation at 60
MHz
Amplitude average
72-points
11,469
213.4us
1.40e-5
4.30e-5
256-points
44,591
829.8us
1.40e-5
4.30e-5
512-points
90,671
1.69ms
1.40e-5
4.30e-5
1024-points
182,831
3.40ms
1.40e-5
4.30e-5
Max. amplitude
More details on Table 4.1.1 in annexes
5.2.2.2
16-bit IIR filter: avr32-uc3 optimized
Concerned file path: /BASIC/FILTERING/dsp16_iir_avr32uc3.c
Algorithm’s size in memory: 1.0 Kbytes (size optimization), 3.1 Kbytes (speed optimization).
Order of the filter: 7
Lowest Error
Lowest cycle
count
Fastest computation at 60
MHz
Amplitude average
72-points
4,332
78.0us
1.90e-5
6.90e-5
256-points
15,006
270.5us
1.70e-5
6.90e-5
512-points
29,854
538.2us
1.70e-5
6.90e-5
1024-points
59,550
1.07ms
1.70e-5
6.90e-5
Max. amplitude
More details on Table 4.1.2 in annexes
5.2.2.3
32-bit IIR filter: generic
Concerned file path: /BASIC/FILTERING/dsp32_iir_generic.c
Algorithm’s size in memory: 400 bytes.
Order of the filter: 7
Lowest Error
Lowest cycle
count
Fastest computation at 60
MHz
Amplitude average
Max. amplitude
72-points
14,517
265.4us
1.50e-9
7.00e-9
256-points
56,471
1.03ms
1.50e-9
7.00e-9
512-points
114,839
2.10ms
1.50e-9
7.00e-9
1024-points
231,575
4.23ms
1.50e-9
7.00e-9
More details on Table 4.2.1 in annexes
16
AVR32718
32076A–AVR32–11/07
AVR32718
5.2.2.4
32-bit IIR filter: avr32-uc3 optimized
Concerned file path: /BASIC/FILTERING/dsp32_iir_avr32uc3.c
Algorithm’s size in memory: 3.0 Kbytes.
Order of the filter: 7
Lowest Error
Lowest cycle
count
Fastest computation at 60
MHz
Amplitude average
72-points
8,859
153.9us
1.50e-9
7.00e-9
256-points
33,333
577.1us
1.50e-9
7.00e-9
512-points
67,381
1.17ms
1.60e-9
7.00e-9
1024-points
135,477
2.34ms
1.60e-9
7.00e-9
Max. amplitude
More details on Table 4.2.2 in annexes
Benchmark results for the 16-bit version (with speed optimization)
The benchmark has been performed on a 72-element input signal.
25000
Cycles
20000
15000
avr32-uc3 optimized
generic
10000
5000
19
17
15
13
11
9
7
5
3
1
0
Number of coefficients
17
32076A–AVR32–11/07
Benchmark results for the 32-bit version (with speed optimization)
The benchmark has been performed on a 72-element input signal.
30000
25000
Cycles
20000
avr32uc3 optimized
15000
generic
10000
5000
19
17
15
13
11
9
7
5
3
1
0
Number of coefficients
Remark: the number of coefficients corresponds to a filter which order is equal to “Number of
coefficients” – 1.
18
AVR32718
32076A–AVR32–11/07
AVR32718
6. Annexes
Optimization
Table 1.1.1 - Benchmark of 16-bit Radix-4 D.I.T. complex FFT: generic
64-point
256-point
1024-point
4096-point
Error
T.F.T.* in SRAM (cycles)
T.F.T.* in FLASH (cycles)
wait-state
wait-state
Ampli- Max.
tude
ampliaveraget u d e
(x10-5) (x10-5)
Algorithm’s size in
memory (bytes)
(code size + T.F.T.*
size)
0 at 30MHz
1 at 60MHz
0 at 30MHz
1 at 60MHz
(0)
6,296 (209.9us)
6,489 (108.2us)
6,640 (221.3us)
7,012 (116.9us)
2.98
15.63
(1)
6,343 (211.4us)
6,538 (109.0us)
6,777 (225.9us)
7,167 (119.5us)
1.58
6.53
1K (844 + 194)
(2)
6,666 (222.2us)
6,959 (116.0us)
7,092 (236.4us)
7,546 (125.8us)
2.98
15.63
1.1K (1032 + 66)
1K (840 + 194)
(3)
6,747 (224.9us)
6,996 (116.6us)
7,101 (236.7us)
7,536 (125.6us)
1.58
6.53
1K (984 + 66)
(0)
33,723 (1124.1us)
34,682 (578.0us)
35,265 (1175.5us)
37,033 (617.2us)
3.02
21.92
1.6K (840 + 770)
(1)
34,041 (1134.7us)
35,003 (583.4us)
35,988 (1199.6us)
37,836 (630.6us)
1.69
8.80
1.6K (844 + 770)
(2)
35,304 (1176.8us)
36,675 (611.3us)
37,194 (1239.8us)
39,294 (654.9us)
3.02
21.92
1.3K (1032 + 258)
(3)
35,,826 (1194.2us)
37,032 (617.2us)
37,419 (1247.3us)
39,453 (657.6us)
1.69
8.80
1.2K (984 + 258)
(0)
169,006 (5.63ms)
173,611 (2.90ms)
175,394 (5.85ms)
183,358 (3.06ms)
3.04
25.01
3.8K (840 + 3K)
(1)
170,795 (5.69ms)
175,404 (2.92ms)
178,863 (5.96ms)
187,161 (3.12ms)
1.67
12.31
3.8K (844 + 3K)
(2)
175,446 (5.85ms)
181,703 (3.03ms)
183,248 (6.11ms)
192,530 (3.21ms)
3.04
25.01
2K (1032 + 1K)
(3)
178,153 (5.94ms)
183,772 (3.06ms)
184,761 (6.16ms)
193,802 (3.23ms)
1.67
12.31
2K (984 + 1K)
(0)
812,321 (27.08ms) 833,820 (13.90ms) 838,147 (27.94ms) 873,235 (14.55ms)
3.09
32.90
12.8K (840 + 12K)
(1)
821,533 (27.38ms) 843,037 (14.05ms) 854,154 (28.47ms) 890,598 (14.84ms)
1.52
14.60
12.8K (844 + 12K)
(2)
838,212 (27.94ms) 866,315 (14.44ms) 869,718 (28.99ms) 910,054 (15.17ms)
3.09
32.90
5K (1032 + 4K)
(3)
851,232 (28.37ms) 876,816 (14.61ms) 877,959 (29.27ms) 917,367 (15.29ms)
1.52
14.60
5K (984 + 4K)
*: Twiddle Factors Table.
(0): Algorithmic optimized for speed.
(1): Algorithmic optimized for accuracy.
(2): Algorithmic optimized for size.
(3): Algorithmic optimized for size and accuracy.
19
32076A–AVR32–11/07
Optimization
Table 1.1.2 - Benchmark of 16-bit Radix-4 D.I.T. complex FFT: avr32-uc3 optimized
64-point
256-point
1024-point
4096-point
Error
T.F.T.* in SRAM (cycles)
T.F.T.* in FLASH (cycles)
wait-state
wait-state
Algorithm’s size in
Amplitude Max.
memory (bytes)
average amplitude (code size + T.F.T.*
(x10-5)
(x10-5)
size)
0 at 30MHz
1 at 60MHz
0 at 30MHz
1 at 60MHz
(0)
2,611 (87.0us)
2,661 (44.4us)
2,753 (91.8us)
2,877 (48.0us)
3.00
10.04
894 (700 + 194)
(1)
2,951 (98.4us)
2,999 (50.0us)
3,097 (103.2us)
3,265 (54.4us)
1.63
6.53
774 (580 + 194)
(2)
2,833 (94.4us)
2,912 (48.5us)
3,027 (100.9us)
3,221 (53.7us)
2.93
10.04
710 (644 + 66)
(3)
3,206 (106.9us)
3,311 (55.2us)
3,458 (115.3us)
3,683 (61.4us)
1.63
6.53
766 (700 + 66)
(0)
13,661 (455.4us)
13,932 (232.2us)
14,306 (476.9us)
14,904 (248.4us)
2.78
13.86
1.4K (700 + 770)
(1)
15,777 (525.9us)
16,033 (267.2us)
16,428 (547.6us)
17,242 (287.4us)
1.68
7.46
1.3K (580 + 770)
(2)
14,651 (488.4us)
15,056 (250.9us)
15,527 (517.6us)
16,442 (274.0us)
2.87
15.82
902 (644 + 258)
(3)
16,916 (563.9us)
17,469 (291.2us)
18,041 (601.4us)
19,152 (319.2us)
1.68
7.46
958 (700 + 258)
(0)
67,671 (2.26ms)
69,027 (1.15ms)
70,355 (2.35ms)
73,059 (1.22ms)
2.84
19.33
3.7K (700 + 3K)
(1)
79,195 (2.64ms)
80,475 (1.34ms)
81,887 (2.73ms)
85,507 (1.43ms)
1.69
10.23
3.6K (580 + 3K)
(2)
71,765 (2.39ms)
73,680 (1.23ms)
75,403 (2.51ms)
79,423 (1.32ms)
2.86
19.33
1.6K (644 + 1K)
(3)
83,906 (2.80ms)
86,635 (1.44ms)
88,560 (2.95ms)
93,629 (1.56ms)
1.69
10.23
1.7K (700 + 1K)
(0)
322,897 (10.76ms)
329,370 (5.49ms)
333,764 (11.13ms)
345,678 (5.76ms)
2.80
23.69
12.7K (700 + 12K)
(1)
381,269 (12.71ms)
387,413 (6.46ms)
392,146 (13.07ms)
407,788 (6.80ms)
1.58
11.84
12.6K (580 + 12K)
(2)
339,439 (11.31ms)
348,176 (5.80ms)
354,159 (11.81ms)
371,396 (6.19ms)
2.81
23.69
4.6K (644 + 4K)
(3)
400,304 (13.34ms)
413,273 (6.89ms)
419,111 (13.97ms)
441,578 (7.36ms)
1.58
11.84
4.7K (700 + 4K)
*: Twiddle Factors Table.
(0): Algorithmic optimized for speed.
(1): Algorithmic optimized for accuracy.
(2): Algorithmic optimized for size.
(3): Algorithmic optimized for size and accuracy.
20
AVR32718
32076A–AVR32–11/07
AVR32718
Optimization
Table 1.2.1 - Benchmark of 32-bit Radix-4 D.I.T. complex FFT: generic
64-point
256-point
1024-point
Error
T.F.T.* in SRAM (cycles)
T.F.T.* in FLASH (cycles)
wait-state
wait-state
Algorithm’s size in
Amplitude Max.
memory (bytes)
average amplitude (code size + T.F.T.*
(x10-9)
(x10-9)
size)
0 at 30MHz
1 at 60MHz
0 at 30MHz
1 at 60MHz
(0)
13,206 (440.2us)
13,509 (225.2us)
13,580 (452.7us)
14,063 (234.4us)
0.70
5.70
(1)
15,323 (510.8us)
15,659 (261.0us)
15,627 (520.9us)
16,113 (268.6us)
0.60
5.70
2.4K (2112 + 392)
(2)
13,622 (454.1us)
14,011 (233.5us)
13,938 (464.6us)
14,497 (241.6us)
0.70
5.70
2.0K (1952 + 136)
(3)
15,714 (523.8us)
16,245 (270.8us)
16,058 (535.3us)
16,805 (280.1us)
0.60
5.70
2.3K (2248 + 136)
(0)
74,297 (2.48ms)
75,940 (1.27ms)
75,992 (2.53ms)
78,445 (1.31ms)
0.50
4.80
3.3K (1816 + 1.5K)
(1)
87,791 (2.93ms)
89,605 (1.49ms)
89,165 (2.97ms)
91,636 (1.53ms)
0.30
4.80
3.6K (2112 + 1.5K)
(2)
76,136 (2.54ms)
78,160 (1.30ms)
77,537 (2.58ms)
80,329 (1.34ms)
0.50
4.80
2.4K (1952 + 520)
(3)
89,543 (2.98ms)
92,446 (1.54ms)
91,058 (3.04ms)
94,978 (1.59ms)
0.30
4.80
2.7K (2248 + 520)
(0)
383,212 (12.77ms)
391,715 (6.53ms)
390,260 (13.01ms)
402,123 (6.70ms)
0.50
8.80
7.8K (1816 + 6K)
(1)
457,571 (15.25ms)
466,815 (7.78ms)
463,279 (15.44ms)
475,223 (7.92ms)
0.30
6.10
8.1K (2112 + 6K)
(2)
390,794 (13.03ms)
400,869 (6.68ms)
396,576 (13.22ms)
409,841 (6.83ms)
0.50
8.80
3.9K (1952 + 2K)
(3)
464,988 (15.50ms)
479,847 (8.00ms)
471,226 (15.71ms)
490,367 (8.17ms)
0.30
6.10
4.2K (2248 + 2K)
2.2K (1816 + 392)
*: Twiddle Factors Table.
(0): Algorithmic optimized for speed.
(1): Algorithmic optimized for accuracy.
(2): Algorithmic optimized for size.
(3): Algorithmic optimized for size and accuracy.
Table 2.1.1 - Benchmark of 16-bit Convolution: generic
Error
Execution time (cycles)
Amplitude average
Max. amplitude
2nd input signal
0 at 30MHz
1 at 60MHz
(x10-5)
(x10-5)
32-points
15,681 (522.7us)
16,344 (272.4us)
1.60
3.90
64-points
23,524 (784.1us)
24,508 (408.5us)
2.00
4.50
128-points
39,204 (1.31ms)
40,830 (680.5us)
1.90
4.50
wait-state
1st input signal
32-point
64-point
128-point
256-point
256-points
70,564 (2.35ms)
73,470 (1.22ms)
1.70
4.10
64-points
57,757 (1.93ms)
60,084 (1.00ms)
1.80
4.70
128-points
86,752 (2.89ms)
90,234 (1.50ms)
1.80
4.40
256-points
144,736 (4.82ms)
150,522 (2.51ms)
1.50
4.80
128-points
221,780 (7.39ms)
230,512 (3.84ms)
1.80
5.30
256-points
333,018 (11.10ms)
346,097 (5.77ms)
1.70
5.00
256-points
869,316 (28.98ms)
903,136 (15.05ms)
1.70
5.40
21
32076A–AVR32–11/07
Table 2.1.2 - Benchmark of 16-bit Convolution: avr32-uc3 optimized
Error
Execution time (cycles)
wait-state
1st input signal
32-point
64-point
128-point
256-point
Amplitude average
Max. amplitude
2nd input signal
0 at 30MHz
1 at 60MHz
(x10-5)
(x10-5)
32-points
5,571 (185.7us)
6,127 (102.1us)
1.60
3.90
64-points
8,248 (274.9us)
9,070 (151.2us)
2.00
4.50
128-points
13,592 (453.1us)
14,944 (249.1us)
1.90
4.50
256-points
24,280 (809.3us)
26,688 (444.8us)
1.70
4.10
64-points
19,087 (636.2us)
20,947 (349.1us)
1.80
4.70
128-points
28,532 (951.1us)
31,308 (521.8us)
1.80
4.40
256-points
47,412 (1.58ms)
52,012 (866.9us)
1.50
4.80
128-points
70,694 (2.36ms)
77,471 (1.29ms)
1.80
5.30
256-points
105,966 (3.53ms)
116,099 (1.93ms)
1.70
5.00
256-points
272,214 (9.07ms)
298,031 (4.97ms)
1.70
5.40
Table 2.2.1 - Benchmark of 32-bit Convolution: generic
Error
Execution time (cycles)
wait-state
1st input signal
32-point
64-point
128-point
256-point
22
Amplitude average
Max. amplitude
2nd input signal
0 at 30MHz
1 at 60MHz
(x10-9)
(x10-9)
32-points
28,359 (945.3us)
29,182 (486.4us)
0.40
1.60
64-points
42,572 (1.42ms)
43,788 (729.8us)
0.40
2.10
128-points
70,988 (2.37ms)
72,990 (1.22ms)
0.30
1.90
256-points
127,820 (4.26ms)
131,390 (2.19ms)
0.40
1.70
64-points
109,179 (3.64ms)
112,334 (1.87ms)
0.40
1.70
128-points
163,968 (5.47ms)
168,678 (2.81ms)
0.50
1.60
256-points
273,536 (9.12ms)
281,350 (4.69ms)
0.60
2.70
128-points
429,026 (14.30ms)
441,458 (7.36ms)
0.40
2.30
256-points
644,074 (21.47ms)
662,677 (11.04ms)
0.40
2.00
256-points
1,701,554 (56.72ms)
1,750,962 (29.18ms)
0.50
3.10
AVR32718
32076A–AVR32–11/07
AVR32718
Table 2.2.2 - Benchmark of 32-bit Convolution: avr32-uc3 optimized
Error
Execution time (cycles)
wait-state
1st input signal
32-point
64-point
128-point
256-point
Amplitude average
Max. amplitude
2nd input signal
0 at 30MHz
1 at 60MHz
(x10-9)
(x10-9)
32-points
13,361 (445.4us)
13,680 (228.0us)
0.60
2.30
64-points
19,958 (665.3us)
20,414 (340.2us)
0.50
2.10
128-points
33,142 (1.10ms)
33,872 (564.5us)
0.50
1.90
256-points
59,510 (1.98ms)
60,784 (1.01ms)
0.50
2.10
64-points
50,501 (1.68ms)
51,624 (860.4us)
0.50
1.70
128-points
75,722 (2.52ms)
77,376 (1.29ms)
0.60
2.40
256-points
126,154 (4.21ms)
128,864 (2.15ms)
0.70
2.70
128-points
196,972 (6.57ms)
201,244 (3.35ms)
0.50
2.30
256-points
295,540 (9.85ms)
301,887 (5.03ms)
0.60
2.30
256-points
778,684 (25.96ms)
795,388 (13.26ms)
0.60
3.10
Number of Taps
Table 3.1.1 - Benchmark of 16-bit FIR Filter: generic
64-point
256-point
512-point
1024-point
Error
I.R.* in SRAM (cycles)
I.R.* in FLASH (cycles)
wait-state
wait-state
Max.
amplitude
(x10-5)
0 at 30MHz
1 at 60MHz
0 at 30MHz
1 at 60MHz
Amplitude
average
(x10-5)
24
7,424 (247.5us)
7,682 (128.0us)
10,704 (356.8us)
12,352 (205.9us)
2.27
9.46
48
5,780 (192.7us)
5,996 (99.9us)
8,517 (283.9us)
9,868 (164.5us)
3.09
15.57
24
41,793 (1.39ms)
43,202 (720.0us)
60,433 (2.01ms)
69,760 (1.16ms)
2.22
9.46
48
70,101 (2.34ms)
72,620 (1.21ms)
103,750 (3.46ms)
120,268 (2.00ms)
5.66
27.20
72
90,921 (3.03ms)
94,262 (1.57ms)
135,691 (4.52ms)
157,528 (2.63ms)
11.25
52.65
100
105,127 (3.50ms)
108,903 (1.82ms)
156,466 (5.22ms)
185,989 (3.10ms)
14.19
63.71
24
87,617 (2.92ms)
90,562 (1.51ms)
126,737 (4.22ms)
146,304 (2.44ms)
2.23
9.46
48
155,861 (5.20ms)
161,452 (2.69ms)
230,726 (7.69ms)
267,468 (4.46ms)
5.78
27.20
72
216,617 (7.22ms)
224,566 (3.74ms)
323,339 (10.78ms)
375,384 (6.26ms)
10.79
52.65
100
276,391 (9.21ms)
286,311 (4.77ms)
411,442 (13.71ms)
489,093 (8.15ms)
14.05
63.71
24
179,265 (5.98ms)
185,282 (3.09ms)
259,345 (8.64ms)
299,392 (4.99ms)
2.21
9.46
48
327,381 (10.91ms)
339,116 (5.65ms)
484,678 (16.16ms)
561,868 (9.36ms)
5.85
27.20
72
468,009 (15.60ms)
485,174 (8.09ms)
698,635 (23.29ms)
811,096 (13.52ms)
10.69
52.65
921,394 (30.71ms)
1,095,301
(18.26ms)
13.99
63.71
100
618,919 (20.63ms)
641,127 (10.69ms)
*: Impulse Response.
23
32076A–AVR32–11/07
Number of Taps
Table 3.1.2 - Benchmark of 16-bit FIR Filter: avr32-uc3 optimized
64-point
256-point
512-point
1024-point
Error
I.R.* in SRAM (cycles)
wait-state
0 at 30MHz
1 at 60MHz
I.R.* in FLASH (cycles)
wait-state
0 at 30MHz
1 at 60MHz
Amplitude
average
(x10-5)
Max.
amplitude
(x10-5)
24
2,439 (81.3us)
2,657 (44.3us)
2,703 (90.1us)
3,054 (50.9us)
2.27
9.46
48
2,115 (70.5us)
2,309 (38.5us)
2,355 (78.5us)
2,670 (44.5us)
3.09
15.57
24
12,712 (423.7us)
13,841 (230.7us)
14,128 (470.9us)
15,966 (266.1us)
2.22
9.46
48
21,604 (720.1us)
23,573 (392.9us)
24,148 (804.9us)
27,390 (456.5us)
5.66
27.20
72
28,192 (939.7us)
30,785 (513.1us)
31,576 (1.05ms)
35,862 (597.7us)
11.25
52.65
100
32,966 (1.10ms)
36,014 (600.2us)
36,966 (1.23ms)
42,015 (700.3us)
14.19
63.71
24
26,408 (880.3us)
28,753 (479.2us)
29,360 (978.7us)
33,182 (553.0us)
2.23
9.46
48
47,588 (1.59ms)
51,925 (865.4us)
53,204 (1.77ms)
60,350 (1.01ms)
5.78
27.20
72
66,464 (2.22ms)
72,577 (1.21ms)
74,456 (2.48ms)
84,566 (1.41ms)
10.79
52.65
100
85,574 (2.85ms)
93,486 (1.56ms)
95,974 (3.20ms)
109,087 (1.82ms)
14.05
63.71
24
53,800 (1.79ms)
58,577 (976.3us)
59,824 (1.99ms)
67,614 (1.13ms)
2.21
9.46
48
99,556 (3.32ms)
108,629 (1.81ms)
111,316 (3.71ms)
126,270 (2.10ms)
5.85
27.20
72
143,008 (4.77ms)
156,161 (2.60ms)
160,216 (5.34ms)
181,974 (3.03ms)
10.69
52.65
100
190,790 (6.36ms)
208,430 (3.47ms)
213,990 (7.13ms)
243,231 (4.05ms)
13.99
63.71
*: Impulse Response.
Number of Taps
Table 3.2.1 - Benchmark of 32-bit FIR Filter: generic
64-point
256-point
512-point
1024-point
Error
I.R.* in SRAM (cycles)
I.R.* in FLASH (cycles)
wait-state
wait-state
Amplitude
average
(x10-9)
Max.
amplitude
(x10-9)
0 at 30MHz
1 at 60MHz
0 at 30MHz
1 at 60MHz
24
13,984 (466.1us)
14,365 (239.4us)
16,608 (553.6us)
18,624 (310.4us)
2.10
12.40
48
11,101 (370.0us)
11,419 (190.3us)
13,311 (443.7us)
14,865 (247.8us)
2.50
14.40
17.40
24
79,073 (2.64ms)
81,181 (1.35ms)
93,985 (3.13ms)
105,408 (1.76ms)
2.30
48
135,518 (4.52ms)
139,291 (2.32ms)
162,688 (5.42ms)
181,713 (3.03ms)
2.70
16.60
72
177,131 (5.90ms)
182,137 (3.04ms)
213,391 (7.11ms)
238,002 (3.97ms)
4.60
31.50
100
187,238 (6.24ms)
194,313 (3.24ms)
241,717 (8.06ms)
264,808 (4.41ms)
4.60
31.90
24
165,857 (5.53ms)
170,269 (2.84ms)
197,153 (6.57ms)
221,120 (3.69ms)
2.60
23.10
48
301,406 (10.05ms)
309,787 (5.16ms)
361,856 (12.06ms)
404,177 (6.74ms)
3.20
23.90
72
422,123 (14.07ms)
434,041 (7.23ms)
508,559 (16.95ms)
567,218 (9.45ms)
5.10
35.30
100
492,390 (16.41ms)
510,985 (8.52ms)
635,701 (21.19ms)
696,424 (11.61ms)
5.10
31.90
24
339,425 (11.31ms)
348,445 (5.81ms)
403,489 (13.45ms)
452,544 (7.54ms)
3.70
28.40
48
633,182 (21.11ms)
650,779 (10.85ms)
760,192 (25.34ms)
849,105 (14.15ms)
4.60
29.30
72
912,107 (30.40ms)
937,849 (15.63ms)
1,098,895
(36.63ms)
1,225,650
(20.43ms)
6.50
35.30
100
1,102,694
(36.76ms)
1,144,329
(19.07ms)
1,423,669
(47.46ms)
1,559,656
(26.00ms)
6.40
33.30
*: Impulse Response.
24
AVR32718
32076A–AVR32–11/07
AVR32718
Number of Taps
Table 3.2.2 - Benchmark of 32-bit FIR Filter: avr32-uc3 optimized
64-point
256-point
512-point
1024-point
Error
I.R.* in SRAM (cycles)
I.R.* in FLASH (cycles)
wait-state
wait-state
Amplitude
average
Max.
amplitude
(x10-9)
(x10-9)
0 at 30MHz
1 at 60MHz
0 at 30MHz
1 at 60MHz
24
6,479 (216.0us)
48
5,132 (171.1us)
6,613 (110.2us)
9,141 (304.7us)
10,918 (182.0us)
2.10
12.40
5,245 (87.4us)
7,305 (243.5us)
8,866 (147.8us)
2.50
14.40
24
36,432 (1.21ms)
37,140 (619.0us)
51,574 (1.72ms)
61,605 (1.03ms)
2.30
12.40
48
62,157 (2.07ms)
63,420 (1.06ms)
88,906 (2.96ms)
107,937 (1.80ms)
2.70
16.60
72
81,114 (2.70ms)
82,788 (1.38ms)
116,446 (3.88ms)
142,173 (2.37ms)
4.60
31.50
100
94,441 (3.15ms)
96,490 (1.61ms)
140,910 (4.70ms)
166,671 (2.78ms)
4.60
31.90
24
76,368 (2.55ms)
77,844 (1.30ms)
108,150 (3.61ms)
129,189 (2.15ms)
2.60
23.10
48
138,189 (4.61ms)
140,988 (2.35ms)
197,706 (6.59ms)
240,033 (4.00ms)
3.20
23.90
35.30
72
193,242 (6.44ms)
197,220 (3.29ms)
277,470 (9.25ms)
338,781 (5.65ms)
5.10
100
248,297 (8.28ms)
253,674 (4.23ms)
370,542 (12.35ms)
438,287 (7.30ms)
5.10
31.90
24
156,240 (5.21ms)
159,252 (2.65ms)
221,302 (7.38ms)
264,357 (4.41ms)
3.70
28.40
48
290,253 (9.68ms)
296,124 (4.94ms)
415,306 (13.84ms)
504,225 (8.40ms)
4.50
29.30
72
417,498 (13.92ms)
426,084 (7.10ms)
599,518 (19.99ms)
731,997 (12.20ms)
6.50
35.30
100
556,009 (18.53ms)
568,042 (9.47ms)
829,806 (27.66ms)
981,519 (16.36ms)
6.40
33.30
*: Impulse Response.
25
32076A–AVR32–11/07
Order of the filter
Table 4.1.1 - Benchmark of 16-bit FIR Filter: generic
72-point
256-point
512-point
1024-point
Error
Execution time (cycles)
wait-state
0 at 30MHz
1 at 60MHz
Amplitude average
Max. amplitude
(x10-5)
(x10-5)
2
5,294 (176.5us)
5,717 (95.3us)
1.50
4.00
7
11,469 (382.3us)
12,802 (213.4us)
1.40
4.30
12
16,319 (544.0us)
18,387 (306.5us)
2.50
7.60
2
19,096 (636.5us)
20,621 (343.7us)
1.50
4.00
4.30
7
44,591 (1.49ms)
49,786 (829.8us)
1.40
12
68,761 (2.29ms)
77,451 (1.29ms)
3.50
8.90
2
38,296 (1.28ms)
41,357 (689.3us)
1.50
4.00
7
90,671 (3.02ms)
101,242 (1.69ms)
1.40
4.30
12
141,721 (4.72ms)
159,627 (2.66ms)
3.70
8.90
2
76,696 (2.56ms)
82,829 (1.38ms)
1.50
4.00
7
182,831 (6.09ms)
204,154 (3.40ms)
1.40
4.30
12
287,641 (9.59ms)
323,979 (5.40ms)
3.90
8.90
Order of the filter
Table 4.1.2 - Benchmark of 16-bit FIR Filter: avr32-uc3 optimized
72-point
256-point
512-point
1024-point
26
Error
Execution time (cycles)
wait-state
0 at 30MHz
1 at 60MHz
Amplitude average
Max. amplitude
(x10-5)
(x10-5)
4.00
2
2,725 (90.8us)
2,871 (47.9us)
1.20
7
4,332 (144.4us)
4,682 (78.0us)
1.90
6.90
12
6,089 (203.0us)
6,534 (108.9us)
3.30
8.40
2
9,167 (305.6us)
9,633 (160.6us)
1.10
4.00
7
15,006 (500.2us)
16,228 (270.5us)
1.70
6.90
12.40
12
22,283 (742.8us)
23,876 (397.9us)
5.10
2
18,127 (604.2us)
19,041 (317.4us)
1.10
4.00
7
29,854 (995.1us)
32,292 (538.2us)
1.70
6.90
12
44,811 (1.49ms)
48,004 (800.1us)
5.60
12.40
2
36,047 (1.20ms)
37,857 (631.0us)
1.10
4.00
7
59,550 (1.99ms)
64,420 (1.07ms)
1.70
6.90
12
89,867 (3.00ms)
96,260 (1.60ms)
5.80
12.40
AVR32718
32076A–AVR32–11/07
AVR32718
Order of the filter
Table 4.2.2 - Benchmark of 32-bit FIR Filter: generic
72-point
256-point
512-point
1024-point
Error
Execution time (cycles)
wait-state
0 at 30MHz
1 at 60MHz
Amplitude average
Max. amplitude
(x10-9)
(x10-9)
2
7,582 (252.7us)
8,072 (134.5us)
1.00
3.80
7
14,517 (483.9us)
15,922 (265.4us)
1.50
7.00
12
19,952 (665.1us)
22,097 (368.3us)
2.60
10.30
2
27,456 (915.2us)
29,232 (487.2us)
0.80
3.80
7
56,471 (1.88ms)
61,922 (1.03ms)
1.50
7.00
12
83,986 (2.80ms)
92,937 (1.55ms)
1.50
10.30
2
55,104 (1.84ms)
58,672 (977.9us)
0.80
3.80
7
114,839 (3.83ms)
125,922 (2.10ms)
1.50
7.00
12
173,074 (5.77ms)
191,497 (3.19ms)
1.40
10.30
2
110,400 (3.68ms)
117,552 (1.96ms)
0.80
3.80
7
231,575 (7.72ms)
253,922 (4.23ms)
1.50
7.00
12
351,250 (11.71ms)
388,617 (6.48ms)
1.30
10.30
Order of the filter
Table 4.2.1 - Benchmark of 32-bit FIR Filter: avr32-uc3 optimized
72-point
256-point
512-point
1024-point
Error
Execution time (cycles)
wait-state
0 at 30MHz
Amplitude average
Max. amplitude
1 at 60MHz
(x10-9)
(x10-9)
3.80
2
5,297 (176.6us)
5,662 (94.4us)
1.00
7
8,859 (295.3us)
9,233 (153.9us)
1.50
7.00
12
11,670 (389.0us)
12,239 (204.0us)
3.00
10.30
2
18,731 (624.4us)
20,014 (333.6us)
0.80
3.80
7
33,333 (1.11ms)
34,625 (577.1us)
1.50
7.00
12
46,816 (1.56ms)
48,855 (814.3us)
1.80
10.30
2
37,419 (1.25ms)
39,982 (666.4us)
0.80
3.80
7
67,381 (2.25ms)
69,953 (1.17ms)
1.60
7.00
12
95,712 (3.19ms)
99,799 (1.66ms)
1.70
10.30
2
74,795 (2.49ms)
79,918 (1.33ms)
0.80
3.80
7
135,477 (4.52ms)
140,609 (2.34ms)
1.60
7.00
12
193,504 (6.45ms)
201,687 (3.36ms)
1.60
10.30
27
32076A–AVR32–11/07
Headquarters
International
Atmel Corporation
2325 Orchard Parkway
San Jose, CA 95131
USA
Tel: 1(408) 441-0311
Fax: 1(408) 487-2600
Atmel Asia
Room 1219
Chinachem Golden Plaza
77 Mody Road Tsimshatsui
East Kowloon
Hong Kong
Tel: (852) 2721-9778
Fax: (852) 2722-1369
Atmel Europe
Le Krebs
8, Rue Jean-Pierre Timbaud
BP 309
78054 Saint-Quentin-enYvelines Cedex
France
Tel: (33) 1-30-60-70-00
Fax: (33) 1-30-60-71-11
Atmel Japan
9F, Tonetsu Shinkawa Bldg.
1-24-8 Shinkawa
Chuo-ku, Tokyo 104-0033
Japan
Tel: (81) 3-3523-3551
Fax: (81) 3-3523-7581
Technical Support
Enter Product Line E-mail
Sales Contact
www.atmel.com/contacts
Product Contact
Web Site
www.atmel.com
Literature Requests
www.atmel.com/literature
Disclaimer: The information in this document is provided in connection with Atmel products. No license, express or implied, by estoppel or otherwise, to any
intellectual property right is granted by this document or in connection with the sale of Atmel products. EXCEPT AS SET FORTH IN ATMEL’S TERMS AND CONDITIONS OF SALE LOCATED ON ATMEL’S WEB SITE, ATMEL ASSUMES NO LIABILITY WHATSOEVER AND DISCLAIMS ANY EXPRESS, IMPLIED OR STATUTORY
WARRANTY RELATING TO ITS PRODUCTS INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTY OF MERCHANTABILITY, FITNESS FOR A PARTICULAR
PURPOSE, OR NON-INFRINGEMENT. IN NO EVENT SHALL ATMEL BE LIABLE FOR ANY DIRECT, INDIRECT, CONSEQUENTIAL, PUNITIVE, SPECIAL OR INCIDENTAL DAMAGES (INCLUDING, WITHOUT LIMITATION, DAMAGES FOR LOSS OF PROFITS, BUSINESS INTERRUPTION, OR LOSS OF INFORMATION) ARISING OUT
OF THE USE OR INABILITY TO USE THIS DOCUMENT, EVEN IF ATMEL HAS BEEN ADVISED OF THE POSSIBILITY OF SUCH DAMAGES. Atmel makes no
representations or warranties with respect to the accuracy or completeness of the contents of this document and reserves the right to make changes to specifications
and product descriptions at any time without notice. Atmel does not make any commitment to update the information contained herein. Unless specifically provided
otherwise, Atmel products are not suitable for, and shall not be used in, automotive applications. Atmel’s products are not intended, authorized, or warranted for use
as components in applications intended to support or sustain life.
© 2007 Atmel Corporation. All rights reserved. Atmel ®, logo and combinations thereof, AVR ® and others are registered trademarks or trademarks of Atmel Corporation or its subsidiaries. Other terms and product names may be trademarks of others.
32076A–AVR32–11/07