Kernel Math Emulation again
Ole.Reinartz at nokia.com
Ole.Reinartz at nokia.com
Fri Dec 3 20:03:41 EST 1999
With soft-float your code generally gets bigger but potentially faster,
because in the case of soft-float the µP can execute the jump instr. to the
math- emulation routines out of line (or the compiler can even inline them).
In the case of kernel math emulation the µP has to handle an exception
everytime a fp- instr. is executed.
On the other hand with kernel math emulation you can potentially run any
pmac- compiled program on your mpc8xx without recompilation. You could even
run a pmac- netscape on it (anyone tried that yet?).
In general you should avoid doing heavy fp- calculation on a µP without HW
support for it. I once tried to run an mpeg3- player on an mpc860 at 50MHz
(compiled with -msoft-float) to decode a 1:30 song, but after 30 minutes I
ran out of patience... :-)
Ole
BTW: I put math-emu into kernel v.2.2.5 after Neil's hint (thank you Neil!)
> -----Original Message-----
> From: EXT Ralf HECKHAUSEN [mailto:RHeckhau at frequentis.com]
> Sent: Freitag, 3. Dezember 1999 9:46
> To: linuxppc-embedded at lists.linuxppc.org
> Subject: Re: Kernel Math Emulation again
>
>
>
> Question:
> Is there any significant difference in execution speed or
> program size in either of the two solutions (compiling with
> soft-float or using the kernel math emulation)?
>
> Ralf
>
> >>> Bill Roman <roman at alerton.com> 12/02 11:32 pm >>>
>
> Neil Blackwood wrote:
> >
> > There is some kernel maths emulation stuff in 2.3.18.
> > You will need to copy it over to which ever version you are
> using. It works.
> > Look for directories math-emu.
> > You will need to change config files and make files as
> well. Its fairly
> > easy. Even I managed it.
>
> Thanks, Neil! I had been very optimistic when I found this
> code in 2.3.18, and
> was disappointed when it turned out it wouldn't compile in
> that context.
> Without your hint, I don't think I would have guessed that
> transplanting it
> into a 2.2 kernel would have worked.
>
> To save everyone else a few hours of grepping and tweaking, here are
> instructions and a patch file to add math emulation to the
> 2.2.13 kernel
> sources from linuxppc.cs.nmt.edu.
>
> 1. Obtain mpc8xx-2.2.13.tgz and linux-2.3.18.tgz from
> ftp://linuxppc.cs.nmt.edu/pub/linuxppc/embedded/. Untar them into two
> separate source trees. (Note: all that's needed from 2.3.18 is the
> math emulator; does anyone know whether that's available separately,
> or could someone split it out and put it on the FTP site?)
>
> 2. Copy the arch/ppc/mathemu directory from the 2.3.18 tree to the
> 2.2.13 tree.
>
> 3. Apply the patch (below). Note that I am building for an
> MBX, so that's
> what .config gets patched for.
>
> 4. make symlinks config dep clean zImage.
>
> The following is the patch.
> --------------------------------------------------------------
> --------------
> diff -ur linux-mpc8xx-2.2.13/.config linux/.config
> --- linux-mpc8xx-2.2.13/.config Fri Oct 22 16:08:42 1999
> +++ linux/.config Wed Dec 1 17:10:01 1999
> @@ -10,9 +10,9 @@
> # CONFIG_PPC64 is not set
> CONFIG_8xx=y
> # CONFIG_MPC821 is not set
> -CONFIG_MPC823=y
> +# CONFIG_MPC823 is not set
> # CONFIG_MPC850 is not set
> -# CONFIG_MPC860 is not set
> +CONFIG_MPC860=y
> # CONFIG_MPC860T is not set
> CONFIG_SERIAL_CONSOLE=y
> # CONFIG_PMAC is not set
> @@ -20,12 +20,13 @@
> # CONFIG_CHRP is not set
> # CONFIG_ALL_PPC is not set
> # CONFIG_APUS is not set
> -# CONFIG_MBX is not set
> +CONFIG_MBX=y
> # CONFIG_RPXLITE is not set
> # CONFIG_RPXCLASSIC is not set
> -CONFIG_BSEIP=y
> +# CONFIG_BSEIP is not set
> # CONFIG_SMP is not set
> CONFIG_MACH_SPECIFIC=y
> +CONFIG_MATH_EMULATION=y
>
> #
> # General setup
> @@ -297,8 +298,7 @@
> # MPC8xx CPM Options
> #
> CONFIG_SCC_ENET=y
> -# CONFIG_SCC1_ENET is not set
> -CONFIG_SCC2_ENET=y
> +CONFIG_SCC1_ENET=y
> # CONFIG_FEC_ENET is not set
> # CONFIG_CPM_IIC is not set
> # CONFIG_UCODE_PATCH is not set
> diff -ur linux-mpc8xx-2.2.13/arch/ppc/Makefile linux/arch/ppc/Makefile
> --- linux-mpc8xx-2.2.13/arch/ppc/Makefile Wed Oct 20 15:33:30 1999
> +++ linux/arch/ppc/Makefile Wed Dec 1 16:16:45 1999
> @@ -41,6 +41,12 @@
> ARCHIVES := arch/ppc/kernel/kernel.o arch/ppc/mm/mm.o
> arch/ppc/lib/lib.o $(ARCHIVES)
> CORE_FILES := arch/ppc/kernel/kernel.o arch/ppc/mm/mm.o
> arch/ppc/lib/lib.o $(CORE_FILES)
>
> +ifdef CONFIG_MATH_EMULATION
> +SUBDIRS += arch/ppc/math-emu
> +ARCHIVES += arch/ppc/math-emu/math-emu.o
> +CORE_FILES += arch/ppc/math-emu/math-emu.o
> +endif
> +
> ifdef CONFIG_XMON
> SUBDIRS += arch/ppc/xmon
> CORE_FILES += arch/ppc/xmon/x.o
> diff -ur linux-mpc8xx-2.2.13/arch/ppc/config.in
> linux/arch/ppc/config.in
> --- linux-mpc8xx-2.2.13/arch/ppc/config.in Fri Oct 22 14:56:10 1999
> +++ linux/arch/ppc/config.in Wed Dec 1 16:29:14 1999
> @@ -37,6 +37,15 @@
> if [ "$CONFIG_ALL_PPC" != "y" ];then
> define_bool CONFIG_MACH_SPECIFIC y
> fi
> +
> +if [ "$CONFIG_8xx" = "y" ]; then
> + bool 'Math emulation' CONFIG_MATH_EMULATION
> +else
> + if [ "$CONFIG_PPC64" != "y" ];then
> + define_bool CONFIG_6xx y
> + fi
> +fi
> +
> endmenu
>
> mainmenu_option next_comment
> diff -ur linux-mpc8xx-2.2.13/arch/ppc/kernel/Makefile
> linux/arch/ppc/kernel/Makefile
> --- linux-mpc8xx-2.2.13/arch/ppc/kernel/Makefile Fri Oct
> 22 15:21:39 1999
> +++ linux/arch/ppc/kernel/Makefile Thu Dec 2 10:34:19 1999
> @@ -30,7 +30,7 @@
> endif
>
> ifeq ($(CONFIG_8xx),y)
> -O_OBJS += m8xx_setup.o softemu8xx.o ppc8xx_pic.o
> +O_OBJS += m8xx_setup.o ppc8xx_pic.o
> ifdef CONFIG_PCI
> O_OBJS += qspan_pci.o
> endif
> diff -ur linux-mpc8xx-2.2.13/arch/ppc/kernel/traps.c
> linux/arch/ppc/kernel/traps.c
> --- linux-mpc8xx-2.2.13/arch/ppc/kernel/traps.c Fri Oct
> 22 14:04:09 1999
> +++ linux/arch/ppc/kernel/traps.c Thu Dec 2 10:32:22 1999
> @@ -230,12 +230,12 @@
> SoftwareEmulation(struct pt_regs *regs)
> {
> int errcode;
> - extern int Soft_emulate_8xx (struct pt_regs *regs);
> + extern int do_mathemu(struct pt_regs *regs);
> extern void print_8xx_pte(struct mm_struct *, unsigned long);
>
> if (user_mode(regs))
> {
> - if ((errcode = Soft_emulate_8xx(regs))) {
> + if ((errcode = do_mathemu(regs))) {
> printk("Software Emulation %s/%d NIP: %lx *NIP: 0x%x code: %x",
> current->comm,current->pid,
> regs->nip, *((uint *)regs->nip), errcode);
> diff -ur linux-mpc8xx-2.2.13/include/asm-ppc/processor.h
> linux/include/asm-ppc/processor.h
> --- linux-mpc8xx-2.2.13/include/asm-ppc/processor.h Fri Oct
> 22 16:10:02 1999
> +++ linux/include/asm-ppc/processor.h Wed Dec 1 16:56:20 1999
> @@ -57,8 +57,33 @@
> #define HID0_BTCD (1<<1) /* Branch target cache
> disable */
>
> /* fpscr settings */
> -#define FPSCR_FX (1<<31)
> -#define FPSCR_FEX (1<<30)
> +#define FPSCR_FX 0x80000000 /* FPU exception summary */
> +#define FPSCR_FEX 0x40000000 /* FPU enabled
> exception summary */
> +#define FPSCR_VX 0x20000000 /* Invalid operation
> summary */
> +#define FPSCR_OX 0x10000000 /* Overflow
> exception summary */
> +#define FPSCR_UX 0x08000000 /* Underflow
> exception summary */
> +#define FPSCR_ZX 0x04000000 /* Zero-devide
> exception summary */
> +#define FPSCR_XX 0x02000000 /* Inexact exception
> summary */
> +#define FPSCR_VXSNAN 0x01000000 /* Invalid op for SNaN */
> +#define FPSCR_VXISI 0x00800000 /* Invalid op for
> Inv - Inv */
> +#define FPSCR_VXIDI 0x00400000 /* Invalid op for
> Inv / Inv */
> +#define FPSCR_VXZDZ 0x00200000 /* Invalid op for
> Zero / Zero */
> +#define FPSCR_VXIMZ 0x00100000 /* Invalid op for
> Inv * Zero */
> +#define FPSCR_VXVC 0x00080000 /* Invalid op for Compare */
> +#define FPSCR_FR 0x00040000 /* Fraction rounded */
> +#define FPSCR_FI 0x00020000 /* Fraction inexact */
> +#define FPSCR_FPRF 0x0001f000 /* FPU Result Flags */
> +#define FPSCR_FPCC 0x0000f000 /* FPU Condition Codes */
> +#define FPSCR_VXSOFT 0x00000400 /* Invalid op for
> software request */
> +#define FPSCR_VXSQRT 0x00000200 /* Invalid op for
> square root */
> +#define FPSCR_VXCVI 0x00000100 /* Invalid op for
> integer convert */
> +#define FPSCR_VE 0x00000080 /* Invalid op
> exception enable */
> +#define FPSCR_OE 0x00000040 /* IEEE overflow
> exception enable */
> +#define FPSCR_UE 0x00000020 /* IEEE underflow
> exception enable */
> +#define FPSCR_ZE 0x00000010 /* IEEE zero divide
> exception enable */
> +#define FPSCR_XE 0x00000008 /* FP inexact
> exception enable */
> +#define FPSCR_NI 0x00000004 /* FPU non IEEE-Mode */
> +#define FPSCR_RN 0x00000003 /* FPU rounding control */
>
> #define _MACH_prep 1
> #define _MACH_Pmac 2 /* pmac or pmac clone (non-chrp) */
>
** Sent via the linuxppc-embedded mail list. See http://lists.linuxppc.org/
More information about the Linuxppc-embedded
mailing list