AM5728: Performance differences against AM437x

Koichiro Tashiro

Part Number: AM5728

Hi,

My customer evaluated AM437x and AM57x performance and got below results.

		[ns]
Operation	AM43xx (1GHz)		AM57xx (1.5GHz)
Logical AND (bool)	1.02		0.78
Logical OR (bool)	1.02		0.66
ADD (16bit binary)	2.70		0.75
Multiply(16bit binary)	2.01		0.99
ADD(32bit binary)	1.02		0.63
Multiply(32bit binary)	2.04		0.96
ADD(32bit floating, single precision)	6.60		0.78
Multiply(32bit floating, single precision)	4.59		0.69
ADD(64bit floating, double precision)	6.60		0.66
Multiply(64bit floating, double precision)	7.14		0.66

These tests are done in 1M loop from Cache, so I think ARM core performance difference is observed.
(Only single core is used for AM57xx)

I know Cortex-A9 inside AM43xx is 2.5DMIPS/MHz and Cortex-A15 inside AM57x is 3.5DMIPS/MHz.
So ~2.1 times differences including operation frequency(1GHz vs 1.5GHz) will be expected.
I think these results are within this range except floating operations which show ~10 times differences.

Do these differences in floating operation come from differences in VFPs ?

Thanks and regards,
Koichiro Tashiro

over 6 years ago

0 jian35385 over 6 years ago

TI__Mastermind 23125 points

Koichiro,
Can you confirm what was the compiler option used for the AM43xx project? That kind of different may be either due to cache or even fundamentally the the floating emu library has been used instead of VFP.
Jian

0 Koichiro Tashiro over 6 years ago in reply to jian35385

TI__Mastermind 22945 points

Hi Jian,

Sorry for the delay.
Customer is using arm-linux-gnueabihf-gcc and below are options used.

ARCH=arm
CXX=arm-linux-gnueabihf-gcc
CXXFLAGS = -DTRG_CORTEX

DEBFLAGS = -g
LDLIBS = -lpthread
CXXFLAGS += -DALIGNATTRIB= -DCDECL= -DHANDLE_WIN32_PRAGMA -DDLL_DECL= -D_REENTRANT= -DHUGEPTR=
CXXFLAGS += -O0 -Wall -Wno-unused
CXXFLAGS += $(INCLUDE)
CXXFLAGS += -DLINUX
CXXFLAGS += $(DEBFLAGS)
CXXFLAGS += -fPIC
LDFLAGS += $(LDLIBS)

Thanks and regards,
Koichiro Tashiro

0 Koichiro Tashiro over 6 years ago in reply to Koichiro Tashiro

TI__Mastermind 22945 points

Hi Jian,
Any news?

0 jian35385 over 6 years ago in reply to Koichiro Tashiro

TI__Mastermind 23125 points

Koichiro,

As a test, can you add the following to rebuild both projects and see if there is a difference?
1. . to enable native hardware support on TI tools, you need to specify '-mfloat-abi=hard' and '-mfpu=vfpv3-d16' options
2. you can turn on verbose mode of the compiler with the "-v" option, that will spit out compiler options that were used.
Thanks
Jian

0 jian35385 over 6 years ago in reply to jian35385

TI__Mastermind 23125 points

Koichiro,
also i wanted to send this past e2e post regarding to how our tools handles floating point.
e2e.ti.com/.../2442347
jian

0 Koichiro Tashiro over 6 years ago in reply to jian35385

TI__Mastermind 22945 points

Hi Jian,

Sorry for my late reply.

Customer is using Linaro toolchain.
Standalone Linaro Toolchain - Linaro GCC 5.3 2016.02 hard-float toolchain

According to your comment in the linked thread, TI Linux SDK points to Linaro tools and it have built-in options.
Does it mean '-mfloat-abi=hard' and '-mfpu=vfpv3-d16' options are enabled by default?

Thanks and regards,
Koichiro Tashiro

0 jian35385 over 6 years ago in reply to Koichiro Tashiro

TI__Mastermind 23125 points

Koichiro,
That is what was documented, about the default built-in, but the benchmark suggested otherwise. To be sure, I would rebuild with explicit options.
Jian

0 Koichiro Tashiro over 6 years ago in reply to jian35385

TI__Mastermind 22945 points

Hi Jian,

Customer confirmed both options are properly used for their build.
Is there any other point customer should check?

Thanks and regards,
Koichiro Tashiro

0 jian35385 over 6 years ago in reply to Koichiro Tashiro

TI__Mastermind 23125 points

Koichiro,
Is it possible run one kernel in debug mode and see if the code actually jumped into floating emulation library? I will also talk to the compiler team to see if there is a quick way to verify.
jian

0 jian35385 over 6 years ago in reply to jian35385

TI__Mastermind 23125 points

Koichiro,
I also got a chance to discuss with the compiler team, he confirmed that we shall check the AM437 gcc build options shall contain "--disable-multilib --with-float=hard --with-fpu=vfpv3-d16" when you enable verbose mode.
I will close the thread to satisfy system audit, but it shall reopen once you replies.
jian

Processors

Processors forum

AM5728: Performance differences against AM437x