GCC Optimization Levels - Seeking Insights

GCC Optimization Levels - Seeking Insights

[Aran Nokan]

[-- Attachment #1: Type: text/plain, Size: 1208 bytes --]

Hello GCC Community,

I recently conducted an experiment where I tested the impact of different
GCC optimization levels on the performance of a code
<https://github.com/leechwort/levenberg-maquardt-example>. I observed that
higher optimization levels didn't necessarily result in faster code
execution.

Is it correct or I have made a mistake? Do we have any other parameters for
optimization?


My make file was as follows:


CC=gcc
> CFLAGS_COMMON=-I.
> LDLAGS=-lm
> DEPS = levmarq.h
> OBJ = main.o levmarq.o
>
> %.o: %.c $(DEPS)
>         $(CC) -c -o $@ $< $(CFLAGS) $(LDLAGS)
>
> main: $(OBJ)
>         gcc -o $@ $^ $(CFLAGS) $(LDLAGS)
>
> # No optimization
> main_no_opt: CFLAGS += -O0
> main_no_opt: $(OBJ)
>         gcc -o $@ $^ $(CFLAGS) $(LDLAGS)
>
> # Basic optimization
> main_opt1: CFLAGS += -O1
> main_opt1: $(OBJ)
>         gcc -o $@ $^ $(CFLAGS) $(LDLAGS)
>
> # Moderate optimization
> main_opt2: CFLAGS += -O2
> main_opt2: $(OBJ)
>         gcc -o $@ $^ $(CFLAGS) $(LDLAGS)
>
> # High optimization
> main_opt3: CFLAGS += -O3
> main_opt3: $(OBJ)
>         gcc -o $@ $^ $(CFLAGS) $(LDLAGS)
>
> # Clean rule
> clean:
>         rm -f *.o main main_no_opt main_opt2 main_opt3
>

Best regards,
Aran

Re: GCC Optimization Levels - Seeking Insights

[Jonathan Wakely]

[-- Attachment #1: Type: text/plain, Size: 1849 bytes --]

On Thu, 21 Dec 2023, 19:33 Aran Nokan via Gcc-help, <gcc-help@gcc.gnu.org>
wrote:

> Hello GCC Community,
>
> I recently conducted an experiment where I tested the impact of different
> GCC optimization levels on the performance of a code
> <https://github.com/leechwort/levenberg-maquardt-example>. I observed that
> higher optimization levels didn't necessarily result in faster code
> execution.
>
> Is it correct or I have made a mistake? Do we have any other parameters for
> optimization?
>

There is no guarantee that optimization make code faster, except that -O0
will invariably be slower.


However, I suspect the problem is that your makefile only builds the .o
objects once, for the first target that needs them. That means they will be
optimized (or not optimized) according to the CFLAGS set when building that
first target.

Unless you do 'make clean' before building each target, you weren't
actually testing what you intend to test.


>
> My make file was as follows:
>
>
> CC=gcc
> > CFLAGS_COMMON=-I.
> > LDLAGS=-lm
> > DEPS = levmarq.h
> > OBJ = main.o levmarq.o
> >
> > %.o: %.c $(DEPS)
> >         $(CC) -c -o $@ $< $(CFLAGS) $(LDLAGS)
> >
> > main: $(OBJ)
> >         gcc -o $@ $^ $(CFLAGS) $(LDLAGS)
> >
> > # No optimization
> > main_no_opt: CFLAGS += -O0
> > main_no_opt: $(OBJ)
> >         gcc -o $@ $^ $(CFLAGS) $(LDLAGS)
> >
> > # Basic optimization
> > main_opt1: CFLAGS += -O1
> > main_opt1: $(OBJ)
> >         gcc -o $@ $^ $(CFLAGS) $(LDLAGS)
> >
> > # Moderate optimization
> > main_opt2: CFLAGS += -O2
> > main_opt2: $(OBJ)
> >         gcc -o $@ $^ $(CFLAGS) $(LDLAGS)
> >
> > # High optimization
> > main_opt3: CFLAGS += -O3
> > main_opt3: $(OBJ)
> >         gcc -o $@ $^ $(CFLAGS) $(LDLAGS)
> >
> > # Clean rule
> > clean:
> >         rm -f *.o main main_no_opt main_opt2 main_opt3
> >
>
> Best regards,
> Aran
>

Re: GCC Optimization Levels - Seeking Insights

[David Brown]

On 21/12/2023 20:57, Jonathan Wakely via Gcc-help wrote:
> On Thu, 21 Dec 2023, 19:33 Aran Nokan via Gcc-help, <gcc-help@gcc.gnu.org>
> wrote:
> 
>> Hello GCC Community,
>>
>> I recently conducted an experiment where I tested the impact of different
>> GCC optimization levels on the performance of a code
>> <https://github.com/leechwort/levenberg-maquardt-example>. I observed that
>> higher optimization levels didn't necessarily result in faster code
>> execution.
>>
>> Is it correct or I have made a mistake? Do we have any other parameters for
>> optimization?
>>
> 
> There is no guarantee that optimization make code faster, except that -O0
> will invariably be slower.
> 
> 
> However, I suspect the problem is that your makefile only builds the .o
> objects once, for the first target that needs them. That means they will be
> optimized (or not optimized) according to the CFLAGS set when building that
> first target.
> 
> Unless you do 'make clean' before building each target, you weren't
> actually testing what you intend to test.
> 
> 

For this test, since there are not many source files, I'd put them all 
on the same gcc line :


default: all
.PHONY: all


OPTS = 0 1 2 3 g fast

PROGS = $(foreach opt, $(OPTS), test.$(opt))
all: $(PROGS)

CC = gcc
CFLAGS_COMMON = -I.
LDLAGS = -lm
DEPS = levmarq.h makefile
SRC = main.c levmarg.c

test.% : $(SRC) $(DEPS)
	gcc -o $@ $(CFLAGS) $(LDLAGS) -O$* $(SRC)

clean:
	rm -f $(PROGS)



(I haven't tried the makefile at all, but it might be a starting point 
for the OP.)


>>
>> My make file was as follows:
>>
>>
>> CC=gcc
>>> CFLAGS_COMMON=-I.
>>> LDLAGS=-lm
>>> DEPS = levmarq.h
>>> OBJ = main.o levmarq.o
>>>
>>> %.o: %.c $(DEPS)
>>>          $(CC) -c -o $@ $< $(CFLAGS) $(LDLAGS)
>>>
>>> main: $(OBJ)
>>>          gcc -o $@ $^ $(CFLAGS) $(LDLAGS)
>>>
>>> # No optimization
>>> main_no_opt: CFLAGS += -O0
>>> main_no_opt: $(OBJ)
>>>          gcc -o $@ $^ $(CFLAGS) $(LDLAGS)
>>>
>>> # Basic optimization
>>> main_opt1: CFLAGS += -O1
>>> main_opt1: $(OBJ)
>>>          gcc -o $@ $^ $(CFLAGS) $(LDLAGS)
>>>
>>> # Moderate optimization
>>> main_opt2: CFLAGS += -O2
>>> main_opt2: $(OBJ)
>>>          gcc -o $@ $^ $(CFLAGS) $(LDLAGS)
>>>
>>> # High optimization
>>> main_opt3: CFLAGS += -O3
>>> main_opt3: $(OBJ)
>>>          gcc -o $@ $^ $(CFLAGS) $(LDLAGS)
>>>
>>> # Clean rule
>>> clean:
>>>          rm -f *.o main main_no_opt main_opt2 main_opt3
>>>
>>
>> Best regards,
>> Aran
>>
> 

Re: GCC Optimization Levels - Seeking Insights

[Aran Nokan]

[-- Attachment #1: Type: text/plain, Size: 3288 bytes --]

Hi,

Thank you. I noticed that it is not working with the following error:

No rule to make target 'test.0',


So I changed it to this:

$(PROGS): test.% : $(SRC) $(DEPS)
>         $(CC) -o $@ $(CFLAGS) $(LDLAGS) -O$* $(SRC)


But now another problem:

make: *** No rule to make target 'levmarg.c', needed by 'test.0'.  Stop.
>

Do you have any idea how to fix it?

Sincerely,
Aran

On Fri, Dec 22, 2023 at 10:26 AM David Brown via Gcc-help <
gcc-help@gcc.gnu.org> wrote:

> On 21/12/2023 20:57, Jonathan Wakely via Gcc-help wrote:
> > On Thu, 21 Dec 2023, 19:33 Aran Nokan via Gcc-help, <
> gcc-help@gcc.gnu.org>
> > wrote:
> >
> >> Hello GCC Community,
> >>
> >> I recently conducted an experiment where I tested the impact of
> different
> >> GCC optimization levels on the performance of a code
> >> <https://github.com/leechwort/levenberg-maquardt-example>. I observed
> that
> >> higher optimization levels didn't necessarily result in faster code
> >> execution.
> >>
> >> Is it correct or I have made a mistake? Do we have any other parameters
> for
> >> optimization?
> >>
> >
> > There is no guarantee that optimization make code faster, except that -O0
> > will invariably be slower.
> >
> >
> > However, I suspect the problem is that your makefile only builds the .o
> > objects once, for the first target that needs them. That means they will
> be
> > optimized (or not optimized) according to the CFLAGS set when building
> that
> > first target.
> >
> > Unless you do 'make clean' before building each target, you weren't
> > actually testing what you intend to test.
> >
> >
>
> For this test, since there are not many source files, I'd put them all
> on the same gcc line :
>
>
> default: all
> .PHONY: all
>
>
> OPTS = 0 1 2 3 g fast
>
> PROGS = $(foreach opt, $(OPTS), test.$(opt))
> all: $(PROGS)
>
> CC = gcc
> CFLAGS_COMMON = -I.
> LDLAGS = -lm
> DEPS = levmarq.h makefile
> SRC = main.c levmarg.c
>
> test.% : $(SRC) $(DEPS)
>         gcc -o $@ $(CFLAGS) $(LDLAGS) -O$* $(SRC)
>
> clean:
>         rm -f $(PROGS)
>
>
>
> (I haven't tried the makefile at all, but it might be a starting point
> for the OP.)
>
>
> >>
> >> My make file was as follows:
> >>
> >>
> >> CC=gcc
> >>> CFLAGS_COMMON=-I.
> >>> LDLAGS=-lm
> >>> DEPS = levmarq.h
> >>> OBJ = main.o levmarq.o
> >>>
> >>> %.o: %.c $(DEPS)
> >>>          $(CC) -c -o $@ $< $(CFLAGS) $(LDLAGS)
> >>>
> >>> main: $(OBJ)
> >>>          gcc -o $@ $^ $(CFLAGS) $(LDLAGS)
> >>>
> >>> # No optimization
> >>> main_no_opt: CFLAGS += -O0
> >>> main_no_opt: $(OBJ)
> >>>          gcc -o $@ $^ $(CFLAGS) $(LDLAGS)
> >>>
> >>> # Basic optimization
> >>> main_opt1: CFLAGS += -O1
> >>> main_opt1: $(OBJ)
> >>>          gcc -o $@ $^ $(CFLAGS) $(LDLAGS)
> >>>
> >>> # Moderate optimization
> >>> main_opt2: CFLAGS += -O2
> >>> main_opt2: $(OBJ)
> >>>          gcc -o $@ $^ $(CFLAGS) $(LDLAGS)
> >>>
> >>> # High optimization
> >>> main_opt3: CFLAGS += -O3
> >>> main_opt3: $(OBJ)
> >>>          gcc -o $@ $^ $(CFLAGS) $(LDLAGS)
> >>>
> >>> # Clean rule
> >>> clean:
> >>>          rm -f *.o main main_no_opt main_opt2 main_opt3
> >>>
> >>
> >> Best regards,
> >> Aran
> >>
> >
>
>
>

Re: GCC Optimization Levels - Seeking Insights

[David Brown]

On 22/12/2023 11:44, Aran Nokan via Gcc-help wrote:
> Hi,
> 
> Thank you. I noticed that it is not working with the following error:
> 
> No rule to make target 'test.0',
> 
> 
> So I changed it to this:
> 
> $(PROGS): test.% : $(SRC) $(DEPS)
>>          $(CC) -o $@ $(CFLAGS) $(LDLAGS) -O$* $(SRC)
> 
> 
> But now another problem:
> 
> make: *** No rule to make target 'levmarg.c', needed by 'test.0'.  Stop.
>>
> 
> Do you have any idea how to fix it?

Yes, that one's easy - fix my spelling mistake!  I think your file is 
called "levmarq.c", while I had written "levmarg.c" (with "g" instead of 
"q").

As I said, I had not tested the makefile at all - it was just a 
suggestion that might help you, and it avoids the problem Jonathan 
pointed out where you were not re-compiling the .o files with the right 
optimisation flags.  If this is just for one-off experimentation, you 
could just use your original makefile but with a "make clean" before 
each "make main_opt?" run.

David



> 
> Sincerely,
> Aran
> 
> On Fri, Dec 22, 2023 at 10:26 AM David Brown via Gcc-help <
> gcc-help@gcc.gnu.org> wrote:
> 
>> On 21/12/2023 20:57, Jonathan Wakely via Gcc-help wrote:
>>> On Thu, 21 Dec 2023, 19:33 Aran Nokan via Gcc-help, <
>> gcc-help@gcc.gnu.org>
>>> wrote:
>>>
>>>> Hello GCC Community,
>>>>
>>>> I recently conducted an experiment where I tested the impact of
>> different
>>>> GCC optimization levels on the performance of a code
>>>> <https://github.com/leechwort/levenberg-maquardt-example>. I observed
>> that
>>>> higher optimization levels didn't necessarily result in faster code
>>>> execution.
>>>>
>>>> Is it correct or I have made a mistake? Do we have any other parameters
>> for
>>>> optimization?
>>>>
>>>
>>> There is no guarantee that optimization make code faster, except that -O0
>>> will invariably be slower.
>>>
>>>
>>> However, I suspect the problem is that your makefile only builds the .o
>>> objects once, for the first target that needs them. That means they will
>> be
>>> optimized (or not optimized) according to the CFLAGS set when building
>> that
>>> first target.
>>>
>>> Unless you do 'make clean' before building each target, you weren't
>>> actually testing what you intend to test.
>>>
>>>
>>
>> For this test, since there are not many source files, I'd put them all
>> on the same gcc line :
>>
>>
>> default: all
>> .PHONY: all
>>
>>
>> OPTS = 0 1 2 3 g fast
>>
>> PROGS = $(foreach opt, $(OPTS), test.$(opt))
>> all: $(PROGS)
>>
>> CC = gcc
>> CFLAGS_COMMON = -I.
>> LDLAGS = -lm
>> DEPS = levmarq.h makefile
>> SRC = main.c levmarg.c
>>
>> test.% : $(SRC) $(DEPS)
>>          gcc -o $@ $(CFLAGS) $(LDLAGS) -O$* $(SRC)
>>
>> clean:
>>          rm -f $(PROGS)
>>
>>
>>
>> (I haven't tried the makefile at all, but it might be a starting point
>> for the OP.)
>>
>>
>>>>
>>>> My make file was as follows:
>>>>
>>>>
>>>> CC=gcc
>>>>> CFLAGS_COMMON=-I.
>>>>> LDLAGS=-lm
>>>>> DEPS = levmarq.h
>>>>> OBJ = main.o levmarq.o
>>>>>
>>>>> %.o: %.c $(DEPS)
>>>>>           $(CC) -c -o $@ $< $(CFLAGS) $(LDLAGS)
>>>>>
>>>>> main: $(OBJ)
>>>>>           gcc -o $@ $^ $(CFLAGS) $(LDLAGS)
>>>>>
>>>>> # No optimization
>>>>> main_no_opt: CFLAGS += -O0
>>>>> main_no_opt: $(OBJ)
>>>>>           gcc -o $@ $^ $(CFLAGS) $(LDLAGS)
>>>>>
>>>>> # Basic optimization
>>>>> main_opt1: CFLAGS += -O1
>>>>> main_opt1: $(OBJ)
>>>>>           gcc -o $@ $^ $(CFLAGS) $(LDLAGS)
>>>>>
>>>>> # Moderate optimization
>>>>> main_opt2: CFLAGS += -O2
>>>>> main_opt2: $(OBJ)
>>>>>           gcc -o $@ $^ $(CFLAGS) $(LDLAGS)
>>>>>
>>>>> # High optimization
>>>>> main_opt3: CFLAGS += -O3
>>>>> main_opt3: $(OBJ)
>>>>>           gcc -o $@ $^ $(CFLAGS) $(LDLAGS)
>>>>>
>>>>> # Clean rule
>>>>> clean:
>>>>>           rm -f *.o main main_no_opt main_opt2 main_opt3
>>>>>
>>>>
>>>> Best regards,
>>>> Aran
>>>>
>>>
>>
>>
>>
> 

Re: GCC Optimization Levels - Seeking Insights

[Jonathan Wakely]

[-- Attachment #1: Type: text/plain, Size: 1475 bytes --]

On Fri, 22 Dec 2023, 10:45 Aran Nokan via Gcc-help, <gcc-help@gcc.gnu.org>
wrote:

> Hi,
>
> Thank you. I noticed that it is not working with the following error:
>
> No rule to make target 'test.0',
>
>
> So I changed it to this:
>
> $(PROGS): test.% : $(SRC) $(DEPS)
> >         $(CC) -o $@ $(CFLAGS) $(LDLAGS) -O$* $(SRC)
>
>
> But now another problem:
>
> make: *** No rule to make target 'levmarg.c', needed by 'test.0'.  Stop.
> >
>
> Do you have any idea how to fix it?
>

You spelled it wrong. levmarq not levmarq

(N.B. This is not a question about using gcc, this is just about how to
write a working makefile)



>
> > >>>
> > >>> main: $(OBJ)
> > >>>          gcc -o $@ $^ $(CFLAGS) $(LDLAGS)
> > >>>
> > >>> # No optimization
> > >>> main_no_opt: CFLAGS += -O0
> > >>> main_no_opt: $(OBJ)
> > >>>          gcc -o $@ $^ $(CFLAGS) $(LDLAGS)
> > >>>
> > >>> # Basic optimization
> > >>> main_opt1: CFLAGS += -O1
> > >>> main_opt1: $(OBJ)
> > >>>          gcc -o $@ $^ $(CFLAGS) $(LDLAGS)
> > >>>
> > >>> # Moderate optimization
> > >>> main_opt2: CFLAGS += -O2
> > >>> main_opt2: $(OBJ)
> > >>>          gcc -o $@ $^ $(CFLAGS) $(LDLAGS)
> > >>>
> > >>> # High optimization
> > >>> main_opt3: CFLAGS += -O3
> > >>> main_opt3: $(OBJ)
> > >>>          gcc -o $@ $^ $(CFLAGS) $(LDLAGS)
> > >>>
> > >>> # Clean rule
> > >>> clean:
> > >>>          rm -f *.o main main_no_opt main_opt2 main_opt3
> > >>>
> > >>
> > >> Best regards,
> > >> Aran
> > >>
> > >
> >
> >
> >
>

Re: GCC Optimization Levels - Seeking Insights

[Jonathan Wakely]

[-- Attachment #1: Type: text/plain, Size: 684 bytes --]

On Fri, 22 Dec 2023, 11:51 Jonathan Wakely, <jwakely.gcc@gmail.com> wrote:

>
>
> On Fri, 22 Dec 2023, 10:45 Aran Nokan via Gcc-help, <gcc-help@gcc.gnu.org>
> wrote:
>
>> Hi,
>>
>> Thank you. I noticed that it is not working with the following error:
>>
>> No rule to make target 'test.0',
>>
>>
>> So I changed it to this:
>>
>> $(PROGS): test.% : $(SRC) $(DEPS)
>> >         $(CC) -o $@ $(CFLAGS) $(LDLAGS) -O$* $(SRC)
>>
>>
>> But now another problem:
>>
>> make: *** No rule to make target 'levmarg.c', needed by 'test.0'.  Stop.
>> >
>>
>> Do you have any idea how to fix it?
>>
>
> You spelled it wrong. levmarq not levmarq
>

Ha, my phone's autocorrect changed the second one.

Re: Re: GCC Optimization Levels - Seeking Insights

[Peter0x44]

Not related to the question directly, but I found the following talk on 
the different optimization settings of GCC to be very informative.

https://youtu.be/GufwdypTfrE

It really helped my understanding, and I also saw how big of a difference 
LTO made.

Best wishes,
Peter D.

Re: GCC Optimization Levels - Seeking Insights

[NightStrike]

[-- Attachment #1: Type: text/plain, Size: 364 bytes --]

On Fri, Dec 22, 2023, 04:25 David Brown via Gcc-help <gcc-help@gcc.gnu.org>
wrote:

>
> For this test, since there are not many source files, I'd put them all
> on the same gcc line :
>
>
> default: all
> .PHONY: all
>
>
> OPTS = 0 1 2 3 g fast
>

It's worth testing with march=native and flto, both provide easy wins if
you can accept any implied limitations.

>

Re: GCC Optimization Levels - Seeking Insights

[Segher Boessenkool]

On Mon, Dec 25, 2023 at 11:21:45AM -0500, NightStrike via Gcc-help wrote:
> On Fri, Dec 22, 2023, 04:25 David Brown via Gcc-help <gcc-help@gcc.gnu.org>
> wrote:
> > OPTS = 0 1 2 3 g fast
> 
> It's worth testing with march=native and flto, both provide easy wins if
> you can accept any implied limitations.

Also -Os is interesting.


Segher