How to multiply 16 bit integer with 64 bit float?

[AeroASM]

I recently changed all my x87 code to SSE, except the following: I have a REAL8, which needs to be multiplied by a 16 bit integer.
The only way I can think to do this is:

Code Select Expand


fild my_integer
fmul my_float
fstp my_float


Is there a way to do this without the FPU?

[raymond]

You could always use the algorithms which were developed in the early days to emulate the x87. :eek

As you showed, using the FPU is very simple. Why don't you want to use the FPU???

Raymond

[AeroASM]

I wrote the whole program using the FPU, then I found out that the FPU is very slow, so I got rid of most of it and replaced it with SSE. I just don't want to rely on the FPU even for one little bit.

[Mark_Larson]

  You can use scalar SSE to do it.

[AeroASM]

Will SSE recognize that one number is an integer and the other is a float?

[raymond]

Converting an integer to one of the float formats is quite easy with CPU instructions. If you have been using the FPU in the past, you must be aware of those formats. Ask if you need to go that route and need help.

Raymond

[Mark_Larson]

Will SSE recognize that one number is an integer and the other is a float?

You can convert the integer to a float and then treat it like a float.

[AeroASM]

How do I convert it?

[raymond]

How do I convert it?

Look at the following description of the various floating point formats.

http://www.ray.masmcode.com/tutorial/fpuchap2.htm#floats

Once you understand it, try to convert an integer by yourself and compare it with the value which would be generated by your assembler. If you do succeed, you will remember how to do it for the rest of your life. If you need more help, I will be glad to provide it.

Raymond

[AeroASM]

Thanks, I have just had a brainwave.

[Farabi]

If I dont made mistake, mul instruction is more fast than the fmul and fimul instruction.

[AeroASM]

But mul doesn't work with floats, and I need to multiply an int by a float to get a float.

[hutch--]

Aero,

Rays suggestion is a good one, you can load the integer into a floating point register and do the rest using the other float value.