I am developing a Goertzel algorithm with this PIC and the floating point math "takes for ever" to to the calculation .
I see the code calls an outside routine ___fsmul and I was wondering if this routine use the HW multiplier available in the 18f252 or it is just a general math code.
How can I verify this ?
Regards,
Zenon
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
To ensure that the computation is being done by the hardware multiplier, your best bet would be to write an assembler subroutine yourself that specifically calls the hardware multiplier. An example of the assembler code should be in the datasheet.
Regards,
William
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
You are right . In the Microchip datasheet for 18f252 there are 2 examples for HW multiplier, but they are for 8 x 8 (signed and unsigned) and also for 16 x 16 (signed and unsigned), but no floating point routine . I am not that good to write a floating point code from there.
I went to the Michochip web site and found AN 575 , but I could not figure out if those examples use HW multiplier.
Regards,
Zenon
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
/* Floating point library in optimized assembly for 8051
* Copyright (c) 2004, Paul Stoffregen, paul@pjrc.com
*
* This program is free software; you can redistribute it and/or
* modify it under the terms of the GNU Library General Public License
* as published by the Free Software Foundation; either version 2
* of the License, or (at your option) any later version.
*
* This library is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU General Public License for more details.
*
* You should have received a copy of the GNU General Public License
* along with this program; if not, write to the Free Software
* Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
*/
#define SDCC_FLOAT_LIB
#include <float.h>
#ifdef FLOAT_ASM_MCS51
// float __fsmul (float a, float b) reentrant
static void dummy(void) _naked
{
_asm
.globl ___fsmul
___fsmul:
// extract the two inputs, placing them into:
// sign exponent mantissa
// ---- -------- --------
// a: sign_a exp_a r4/r3/r2
// b: sign_b exp_b r7/r6/r5
lcall fsgetargs
// first check if either input is zero
cjne r4, #0, 00002$
00001$:
ljmp fs_return_zero
00002$:
mov a, r7
jz 00001$
// compute final sign bit
jnb sign_b, 00003$
cpl sign_a
00003$:
// add the exponents
mov a, exp_a
add a, exp_b
add a, #130
mov exp_a, a
// now we need to multipy r4/r3/r2 * r7/r6/r5
// ------------------------------------------
// r2 * r5 << 0
// r3 * r5 + r2 * r6 << 8
// r4 * r5 + r3 * r6 + r2 * r7 << 16
// r4 * r6 + r3 * r7 << 24
// r4 * r7 << 32
//
// This adds quite a bit of code, but it is a LOT faster
// than three calls to __mululong...
// output goes into r4/r3/r2/r1/r0/xx
mov a, r2
mov b, r5
mul ab // r2 * r5
// discard lowest 8 bits
mov r0, b
// range 0-FE
mov a, r2
mov b, r6
mul ab // r2 * r6
add a, r0
mov r0, a
clr a
addc a, b
mov r1, a
// range 0-FEFF
mov a, r3
mov b, r5
mul ab // r3 * r5
add a, r0
// discard lowest 8 bits
mov a, r1
addc a, b
mov r1, a
clr a
rlc a
xch a, r2
// range 0-1FD
mov b, r7
mul ab // r2 * r7
add a, r1
mov r1, a
mov a, r2
addc a, b
mov r2, a
// range 0-FFFE
mov a, r3
mov r0, a
mov b, r6
mul ab // r3 * r6
add a, r1
mov r1, a
mov a, r2
addc a, b
mov r2, a
clr a
rlc a
mov r3, a
// range 0-1FDFF
mov a, r4
mov b, r5
mul ab // r4 * r5
add a, r1
mov r1, a
mov a, r2
addc a, b
mov r2, a
clr a
addc a, r3
mov r3, a
// range 0-2FC00
mov a, r0 // r3
mov b, r7
mul ab // r3 * r7
add a, r2
mov r2, a
mov a, r3
addc a, b
mov r3, a
clr a
rlc a
xch a, r4
// range 0-100FD00
mov r5, a
mov b, r6
mul ab // r4 * r6
add a, r2
mov r2, a
mov a, r3
addc a, b
mov r3, a
clr a
addc a, r4
mov r4, a
// range 0-1FEFE00
mov a, r5 // r4
mov b, r7
mul ab // r4 * r7
add a, r3
mov r3, a
mov a, r4
addc a, b
mov r4, a
// range 40000000-FFFFFE00
jb acc.7, 00010$
lcall fs_normalize_a
00010$:
ljmp fs_round_and_return
_endasm;
}
#else
/*
** libgcc support for software floating point.
** Copyright (C) 1991 by Pipeline Associates, Inc. All rights reserved.
** Permission is granted to do *anything* you want with this file,
** commercial or otherwise, provided this message remains intact. So there!
** I would appreciate receiving any updates/patches/changes that anyone
** makes, and am willing to be the repository for said changes (am I
** making a big mistake?).
**
** Pat Wood
** Pipeline Associates, Inc.
** pipeline!phw@motown.com or
** sun!pipeline!phw or
** uunet!motown!pipeline!phw
*/
/* (c)2000/2001: hacked a little by johan.knol@iduna.nl for sdcc */
union float_long
{
float f;
unsigned long l;
};
/* multiply two floats */
float __fsmul (float a1, float a2) {
volatile union float_long fl1, fl2;
volatile unsigned long result;
volatile int exp;
char sign;
If you (or anyone:) is interested in what it takes to implement assembler floating point routines for a specific target then have a look at Paul Stoffregen's commits in December 2004 for the mcs51.
Greetings,
Frieder
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Hi,
I am developing a Goertzel algorithm with this PIC and the floating point math "takes for ever" to to the calculation .
I see the code calls an outside routine ___fsmul and I was wondering if this routine use the HW multiplier available in the 18f252 or it is just a general math code.
How can I verify this ?
Regards,
Zenon
Hi Zenon.
To ensure that the computation is being done by the hardware multiplier, your best bet would be to write an assembler subroutine yourself that specifically calls the hardware multiplier. An example of the assembler code should be in the datasheet.
Regards,
William
William,
You are right . In the Microchip datasheet for 18f252 there are 2 examples for HW multiplier, but they are for 8 x 8 (signed and unsigned) and also for 16 x 16 (signed and unsigned), but no floating point routine . I am not that good to write a floating point code from there.
I went to the Michochip web site and found AN 575 , but I could not figure out if those examples use HW multiplier.
Regards,
Zenon
I found this site http://www.koders.com/c/fid2AF565F87D85A38001342EA97AA3ADAA28105486.aspx?s=__fsmul that shows the code for floating point multiply routine and it looks is for 8051 CPU and not PIC.
Can any one comment on this ?
/* Floating point library in optimized assembly for 8051
* Copyright (c) 2004, Paul Stoffregen, paul@pjrc.com
*
* This program is free software; you can redistribute it and/or
* modify it under the terms of the GNU Library General Public License
* as published by the Free Software Foundation; either version 2
* of the License, or (at your option) any later version.
*
* This library is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU General Public License for more details.
*
* You should have received a copy of the GNU General Public License
* along with this program; if not, write to the Free Software
* Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
*/
#define SDCC_FLOAT_LIB
#include <float.h>
#ifdef FLOAT_ASM_MCS51
// float __fsmul (float a, float b) reentrant
static void dummy(void) _naked
{
_asm
.globl ___fsmul
___fsmul:
// extract the two inputs, placing them into:
// sign exponent mantissa
// ---- -------- --------
// a: sign_a exp_a r4/r3/r2
// b: sign_b exp_b r7/r6/r5
lcall fsgetargs
// first check if either input is zero
cjne r4, #0, 00002$
00001$:
ljmp fs_return_zero
00002$:
mov a, r7
jz 00001$
// compute final sign bit
jnb sign_b, 00003$
cpl sign_a
00003$:
// add the exponents
mov a, exp_a
add a, exp_b
add a, #130
mov exp_a, a
// now we need to multipy r4/r3/r2 * r7/r6/r5
// ------------------------------------------
// r2 * r5 << 0
// r3 * r5 + r2 * r6 << 8
// r4 * r5 + r3 * r6 + r2 * r7 << 16
// r4 * r6 + r3 * r7 << 24
// r4 * r7 << 32
//
// This adds quite a bit of code, but it is a LOT faster
// than three calls to __mululong...
// output goes into r4/r3/r2/r1/r0/xx
mov a, r2
mov b, r5
mul ab // r2 * r5
// discard lowest 8 bits
mov r0, b
// range 0-FE
mov a, r2
mov b, r6
mul ab // r2 * r6
add a, r0
mov r0, a
clr a
addc a, b
mov r1, a
// range 0-FEFF
mov a, r3
mov b, r5
mul ab // r3 * r5
add a, r0
// discard lowest 8 bits
mov a, r1
addc a, b
mov r1, a
clr a
rlc a
xch a, r2
// range 0-1FD
mov b, r7
mul ab // r2 * r7
add a, r1
mov r1, a
mov a, r2
addc a, b
mov r2, a
// range 0-FFFE
mov a, r3
mov r0, a
mov b, r6
mul ab // r3 * r6
add a, r1
mov r1, a
mov a, r2
addc a, b
mov r2, a
clr a
rlc a
mov r3, a
// range 0-1FDFF
mov a, r4
mov b, r5
mul ab // r4 * r5
add a, r1
mov r1, a
mov a, r2
addc a, b
mov r2, a
clr a
addc a, r3
mov r3, a
// range 0-2FC00
mov a, r0 // r3
mov b, r7
mul ab // r3 * r7
add a, r2
mov r2, a
mov a, r3
addc a, b
mov r3, a
clr a
rlc a
xch a, r4
// range 0-100FD00
mov r5, a
mov b, r6
mul ab // r4 * r6
add a, r2
mov r2, a
mov a, r3
addc a, b
mov r3, a
clr a
addc a, r4
mov r4, a
// range 0-1FEFE00
mov a, r5 // r4
mov b, r7
mul ab // r4 * r7
add a, r3
mov r3, a
mov a, r4
addc a, b
mov r4, a
// range 40000000-FFFFFE00
jb acc.7, 00010$
lcall fs_normalize_a
00010$:
ljmp fs_round_and_return
_endasm;
}
#else
/*
** libgcc support for software floating point.
** Copyright (C) 1991 by Pipeline Associates, Inc. All rights reserved.
** Permission is granted to do *anything* you want with this file,
** commercial or otherwise, provided this message remains intact. So there!
** I would appreciate receiving any updates/patches/changes that anyone
** makes, and am willing to be the repository for said changes (am I
** making a big mistake?).
**
** Pat Wood
** Pipeline Associates, Inc.
** pipeline!phw@motown.com or
** sun!pipeline!phw or
** uunet!motown!pipeline!phw
*/
/* (c)2000/2001: hacked a little by johan.knol@iduna.nl for sdcc */
union float_long
{
float f;
unsigned long l;
};
/* multiply two floats */
float __fsmul (float a1, float a2) {
volatile union float_long fl1, fl2;
volatile unsigned long result;
volatile int exp;
char sign;
fl1.f = a1;
fl2.f = a2;
if (!fl1.l || !fl2.l)
return (0);
/* compute sign and exponent */
sign = SIGN (fl1.l) ^ SIGN (fl2.l);
exp = EXP (fl1.l) - EXCESS;
exp += EXP (fl2.l);
fl1.l = MANT (fl1.l);
fl2.l = MANT (fl2.l);
/* the multiply is done as one 16x16 multiply and two 16x8 multiples */
result = (fl1.l >> 8) * (fl2.l >> 8);
result += ((fl1.l & (unsigned long) 0xFF) * (fl2.l >> 8)) >> 8;
result += ((fl2.l & (unsigned long) 0xFF) * (fl1.l >> 8)) >> 8;
if (result & SIGNBIT)
{
/* round */
result += 0x80;
result >>= 8;
}
else
{
/* round */
result += 0x40;
result >>= 7;
exp--;
}
result &= ~HIDDEN;
/* pack up and go home */
fl1.l = PACK (sign ? SIGNBIT : 0 , (unsigned long)exp, result);
return (fl1.f);
}
#endif
Hi Zenon,
the site you found is an old mirror of SDCC's outdated CVS repository. SDCC has switched to SVN (subversion) quite a while ago.
You'll find a more up to date file either in SDCC's source distribution or at:
http://sdcc.svn.sourceforge.net/viewvc/sdcc/trunk/sdcc/device/lib/_fsmul.c?view=log
The first part of the file contains hand optimized assembler code for the 8051 port and the second part a generic C version.
Either this file or one of the corresponding files:
http://sdcc.svn.sourceforge.net/viewvc/sdcc/trunk/sdcc/device/lib/pic/libsdcc/fsmul.c?view=log
http://sdcc.svn.sourceforge.net/viewvc/sdcc/trunk/sdcc/device/lib/pic16/libsdcc/float/fsmul.c?view=log
is likely being used for your PIC (I'm not familiar with later PICs).
If you (or anyone:) is interested in what it takes to implement assembler floating point routines for a specific target then have a look at Paul Stoffregen's commits in December 2004 for the mcs51.
Greetings,
Frieder