Teensy 3.0

[pito]

..not sure your measurement does reflect the reality

 //time elasped = 29721 micros = teensy
 //time elasped = 47436 micros = Uno

A typical 32bit float sin() on an CM3 takes ~1050cycles = ~22usec @48MHz so it seems you have to get something like 9000 micros max...

CM3 none FPU :
fZ = fX * fY; // 41 cycles
fZ = sqrt(fY); // 624 cycles
fZ = sin(1.23); // 1017 cycles

CM4 with FPU:
fZ = fX * fY; // 6 cycles
fZ = sqrt(fY); // 20 cycles
fZ = sin(1.23); // 124 cycles

http://www.micromouseonline.com/2011/10/26/stm32f4-the-first-taste-of-speed/?doing_wp_cron=1350295560.3099548816680908203125#axzz29MQwGQCw

[cyclegadget]

  I can only say that my code saves the results to an array which is printed after the math has been timed. Could the saving to the array cause the time difference?

[pito]

.. a clock2clock comparision says teensy shall be ~3times faster than Uno (@16MHz) and teensy is 32bit CM3, so a 32bit fp sin() cannot be "only" 1.6x faster than Uno.. saving to an array cannot create such overhead, indeed..

[Jantje]

.. a clock2clock comparision says teensy shall be ~3times faster than Uno (@16MHz) and teensy is 32bit CM3, so a 32bit fp sin() cannot be "only" 1.6x faster than Uno.. saving to an array cannot create such overhead, indeed..

My 2 cents: A possible explanation may be that the compile options for teensy are not similar to those of UNO.
Best regards
Jantje

[pito]

This is with STM32F100 CM3 @48MHz (Teensy compatible I think):

Code: [Select]


timer = millis;
         for (i=0;i<400;i++) {
           sinres[i]=sinf((float)i);
         }
timer = millis - timer;
printf("\rElapsed time float sin 400x into array: %u millis\n", timer);
         
timer = millis;
         for (i=0;i<400;i++) {
           sinresd[i]=sinl((long double)i);
         }
timer = millis - timer;
printf("\rElapsed time double sin 400x into array: %u millis\n", timer);


Elapsed time float sin 400x into array: 8 millis
Elapsed time double sin 400x into array: 19 millis

[pjrc]

There are indeed some complex things going on with this test.

For example, this takes 138 us:

Code: [Select]


 for (int i = 0; i < 3; i++)
 {
   sinanswers[i] = sin(i);
 }
 time2 = micros();


But this takes takes 229 us.... almost twice as long, just because the input is offset by 400.  Clearly sin()'s execution time is not constant.

Code: [Select]


 for (int i = 0; i < 3; i++)
 {
   sinanswers[i] = sin(i+400);
 }
 time2 = micros();


I suspected the slowness was due to computing double precision.  But I tried changing sin() to sinf(), and amazingly sinf() takes MUCH longer.  Clearly newlib or libgcc is not optimized very well, or some settings aren't quite right.  I need to dig into that......

[pito]

And your test with 1000x (actually 500x) sin cos tan (STM32F100 CM3 @48MHz):

Code: [Select]


timer = millis;
         for (i=0;i<500;i++) {
           fsi[i]=sinf((float)i);
           fco[i]=cosf((float)i);
           fta[i]=tanf((float)i);
         }
timer = millis - timer;
printf("\rElapsed time float sin cos tan 500x into array: %u millis\n", timer);

Elapsed time float sin cos tan 500x into array: 31 millis

Such big arrays do not fit into my 8kB RAM so double it for 1000x (=62 millis, yours is 278 ms). Double it again for a double precision fp result.
p.

[fat16lib]

First the Teensy Kinetis CPU does not have hardware floating point.  Hardware floating point is optional for Cortex M4.  Only the high end K20 processors have floating point http://www.freescale.com/files/microcontrollers/doc/fact_sheet/KNTSK20FMLYFS.pdf.

newlib math function are really old C functions.

newlib execution times depend on the value of the arguments.

Here are two examples for 32-bit sine:

Code: [Select]

float sinf(float);

I ran this code

Code: [Select]


float sinanswers[401];
float sinarg[401];
 for (int i = 0; i < 400; i++) {
   sinarg[i] = factor*i;
 }
 time1 = micros();
 for (int i = 0; i < 400; i++) {
   sinanswers[i] = sinf(sinarg[i]);
 }
 time2 = micros();

If factor is 0.01 so the range is from 0.0 - 4.0

time elapsed = 17110 micros

If factor is 1.0 so the range is 0.0 - 400.0

time elapsed = 105353 micros

The algorithms for 64-bit double are totally different than for 32-bit float.

Much of this dates back to work in the 1980s on BSD Unix at UC Berkeley.  I was at UCB when BSD Unix was developed.

Bill Joy was a key developer of BSD and used it at Sun Microsystems as the base for Solaris.

[DuaneB]

Hi,
   I assume that most of us will use fixed point maths, but for those that have a reason to use float and double, is there an alternative implementation that can be included at compile time or some other work around that provide more recent and faster implementations ?

  Duane B

rcarduino.blogspot.com

[westfw]

avr-gcc has floating point algorithms that have been carefully optimized for the AVR architecture.
arm-gcc using newlib presumably has generic algorithms...

[fat16lib]

Thanks to Paul's help, I have first results for SdFat with large reads and writes.

Writes are slow since I am not using the full TX fifo yet.

Here are results for 4096 byte writes and reads using the SdFat bench.ino example.

Type is FAT16
File size 5MB
Buffer size 4096 bytes
Starting write test.  Please wait up to a minute
Write 723.07 KB/sec
Maximum latency: 21609 usec, Minimum Latency: 4929 usec, Avg Latency: 5625 usec

Starting read test.  Please wait up to a minute
Read 1255.56 KB/sec
Maximum latency: 3874 usec, Minimum Latency: 3204 usec, Avg Latency: 3260 usec

I think I know how to speedup writes by using all four bytes of the SPI fifo.

DMA may be required to achieve maximum speed since using the fifo has a bit more overhead than I first assumed.

[Jantje]

fat16lib
I don't know much about sdcards and read and write speeds but those read figures look impressive to me.
Best regards
Jantje

[MichaelMeissner]

What does the S in sin.S stand for? I googled it but it is to close to sin to find something relevant quickly.

The GCC compiler passes .S files through the C preprocessor, so that you can use #ifdef and #define within assembly files (because of the #ifdef, you can have one .S file that has code for several different targets, and the defines are set based on the -m<xxx> options on the command line).  If the file is .s (lowercase), it is passed directly to the assembler and does not go through the preprocessor.

[fat16lib]

Teensy 3.0 is great.  The hardware is a little gem and the way Paul did the software makes development really easy.

I now have optimized native SPI running in programmed I/O mode and performance is really good.  Here are my latest results for SdFat.

Free RAM: 4219
Type is FAT16
File size 10MB
Buffer size 8192 bytes
Starting write test.  Please wait up to a minute
Write 1801.74 KB/sec
Maximum latency: 80495 usec, Minimum Latency: 4398 usec, Avg Latency: 4543 usec

Starting read test.  Please wait up to a minute
Read 2017.82 KB/sec
Maximum latency: 4473 usec, Minimum Latency: 4048 usec, Avg Latency: 4057 usec

This is with the CPU overclocked at 96 MHz.  I used an industrial SD designed for embedded use.

[TonyD]

Great board.

Is there any news when we get buy one? Are there any European distributors stocking it?