[Ffmpeg-devel] int vs. float: Hard Numbers

[Rich Felker]

On Fri, May 20, 2005 at 12:13:24PM -0600, Mike Melanson wrote:
> Hi,
> 	Persuant to the discussion about whether int or float ops were more 
> efficient on today's production CPUs, I performed some empirical tests.
> 
> 	Tests written using gcc+nasm and clocked using the Intel timestamp 
> counter register (TSC) found in Pentium+ CPUs. I compared additions and 
> multiplications. Results:
> 
> integer_adder() (10 adds) returned 50, 36 cycles used
> float_adder() (10 adds) returned 50.000000, 36 cycles used
> integer_mult() (10 mults) returned 9765625, 115 cycles used
> float_mult() (10 mults) returned 9765625.000000, 36 cycles used
> 
> Interesting notes:
> * the "float" routines are all actually 64-bit double math; I just 
> didn't name them correctly
> * the read TSC instruction + other overhead between the rdtsc markers 
> was 26 cycles; thus, for 3/4 tests, each instruction actually required 1 
> clock cycle
> * the cycle count for integer_adder() floated around (36 was the lower 
> limit) but the cycle counts for the others are highly consistent
> * the integer_mult() function had a few more instructions between the 
> markers (2 extra movs), but I do not think that entirely accounts for 
> the discrepancy
> 
> 	If anyone wants the code for these tests, let me know.
> 
> 	Felker: I await your typical analysis based on these results 
> 	("OMG!!1! Intel sUx!!"). These tests were not conducted specifically on an 
> Intel-branded chip, though it is a P3-class chip.

Yes, Intel sucks. Try it on AMD. FYI, test results I recently saw on
#mplayerdev or one of the mailing lists (can't remember which right
off) showed P4 taking 4x as many cycles to perform common codec
operations as an Athlon took.

Rich