using halfs?

[daydreamer]

to save space,could I have lots of data in halfs instead of real4's?
are there already conversion macros/proc between them? or I need to do bitfiddling myself?

[jj2007]

Search the forum for REAL2, there are many posts.

[InfiniteLoop]

F16C is 16-bit floating point.

VCVTPH2PS ymm1, xmm2/m128

[jj2007]

Nice find, InfiniteLoop

I've put together a simple example (using plain Masm32 SDK code, see attached source), but I can't test it on my old cpu...

  vcvtps2ph half, xmm0, 0       ; 4 x single to half-precision; round to nearest even
  vcvtph2ps xmm0, xmm1          ; 4 x half to single-precision

[daydreamer]

Thanks infiniteloop and jj
I try 16 bit fixed point storage ->  fimul constant also

[jj2007]

Did my proggie work for you, daydreamer?

[hutch--]

Seems to run here OK.

test 111.125000
test 222.250000
test 333.250000
test 444.500000
ok?

[LiaoMi]

Nice find, InfiniteLoop

I've put together a simple example (using plain Masm32 SDK code, see attached source), but I can't test it on my old cpu...

  vcvtps2ph half, xmm0, 0       ; 4 x single to half-precision; round to nearest even
  vcvtph2ps xmm0, xmm1          ; 4 x half to single-precision

test 111.125000
test 222.250000
test 333.250000
test 444.500000
ok?

[daydreamer]

Did my proggie work for you, daydreamer?

it works

[daydreamer]

If I have lots of 3d object data, half's would be good to have smaller size data
But even word data = in millimeters might be good alternative

[HSE]

... but I can't test it on my old cpu...

Wrong, JJ

You can use Intel SDE (I don't have AVX, pretty old machine):

Code Select Expand

test 111.125000
test 222.250000
test 333.250000
test 444.500000
ok?

[jj2007]

You can use Intel SDE

Never seen. Interesting, how does it work? I saw it's not such a fat download. Is it freeware? Does it really emulate the processor, in the sense: utterly slow but it works correctly?

[daydreamer]

... but I can't test it on my old cpu...

Wrong, JJ

You can use Intel SDE (I don't have AVX, pretty old machine):

Code Select Expand

test 111.125000
test 222.250000
test 333.250000
test 444.500000
ok?

Nice
But isn't it bigger chance old gpu have half caps,where I seen its used many years ago,and maybe some shader capable of conversion,or shader compiler running on cpu have caps. To convert half's <>floats?

[HSE]

how does it work? I saw it's not such a fat download. Is it freeware?

I don't remember if requiere a free subscription, but is free for our use. Nidud posted a test (or something) a year ago.

There is a batch that open a command window. From the prompt you have a AVX machine 

Also work a little batch:

Code Select Expand

path\sde -- halfprecision.exe

[HSE]

But isn't it bigger chance old gpu have half caps,where I seen its used many years ago,and maybe some shader capable of conversion,or shader compiler running on cpu have caps. To convert half's <>floats?

Have you writed some code to show how that can be maked? That could be interesting