Benchmarks

Performance benchmarks

The benchmarks below are shown in different configurations:

• ro is with rounding and overflow detection enabled.
• nn is -DFIXMATH_NO_OVERFLOW -DFIXMATH_NO_ROUNDING, ie. both disabled.

The number after the configuration means the wordwidth setting, which does not affect accuracy but is selected to be suitable for the platform:

• 64 is default
• 32 is with -DFIXMATH_NO_64BIT
• 08 is with -DFIXMATH_OPTIMIZE_8BIT

The performance tests are performed on a quite small number of testcases, because of the memory needed to store correct answers. The average number may not be representative of the average value in a real application, but the minimum and maximum values likely are quite accurate.

The performance is compared to compiler's own 32-bit float implementation.

ARM Cortex M3

Using simulator qemu-system-arm version 0.15.50 and compiler Sourcery G++ Lite 2011.03-42. The values are the number of instructions executed, including function call and storing result to memory. Most instructions on the platform take 1 cycle to execute, with some taking 2 cycles.

Compiler options: -mcpu=cortex-m3 -mthumb -O2

Operation		min	avg	max	configuration	code size (bytes)
add	fix16_add	9	9	11	ro64	26
	fix16_add	2	2	2	nn64	~2 (inlined)
	float add	22	50	63		352
sub	fix16_sub	9	10	11	ro64	26
	fix16_sub	2	2	2	nn64	~2 (inlined)
	float sub	23	51	64		356
mul	fix16_mul	12	17	20	ro64	76
	fix16_mul	8	8	8	nn64	12
	float mul	23	31	36		360
div	fix16_div	37	54	72	ro64	182
	fix16_div	37	47	62	nn64	166
	float div	23	123	152		310
sqrt	fix16_sqrt	142	169	225	ro64	126
	fix16_sqrt	139	166	222	nn64	120
	float sqrtf	361	361	361		354

Atmel ATMega 128

Using simulator simulavrxx git revision as of 2012-01-05 and compiler avr-gcc 4.5.3. The values are the number of cpu cycles, including function call and storing result to memory.

Compiler options: -mmcu=atmega128 -O2

Operation		min	avg	max	configuration	code size (bytes)
add	fix16_add	49	52	60	ro08	74
	fix16_add	14	14	14	nn08	~14 (inlined)
	float add	606	982	1262		90 (+ 800 shared fp stuff)
sub	fix16_sub	49	53	60	ro08	74
	fix16_sub	14	14	14	nn08	~14 (inlined)
	float sub	612	988	1231		98
mul	fix16_mul	294	344	453	ro08	818
	fix16_mul	205	257	379	nn08	696
	float mul	466	1742	2112		500
div	fix16_div	171	752	1329	ro08	414
	fix16_div	162	738	1307	nn08	382
	float div	438	1293	1370		348
sqrt	fix16_sqrt	605	714	934	ro08	386
	fix16_sqrt	599	707	925	nn08	368
	float sqrtf	486	493	497		144