Archives Discussions

[malcolm3141]

Hi Guys,

The IL spec details an IL instruction called IFB (integer first bit) with the following options:

ifb_lo (search from lsb)

ifb_high (search from msb)

ifb_signed (search from msb as signed)

The CAL IL compiler doesn't accept any of these instructions. It does however accept the instruction 'ffb' which is equivalent to 'ifb_lo'.

Does anyone know if it is possible to get the equivalent to ifb_high with the current compiler?

Malcolm

[MicahVillmow]

malcolm,
The text syntax is a bug that has been fixed internally. There currently is no plan to expose all IL instructions at the OpenCL level at this time.

[malcolm3141]

Hi Micah,

Thanks for the reply. Please note, I'm specifically talking about CAL IL compiler, not OpenCL.

Just to clarify, are you saying that the CAL IL compiler some time in the future will accept the specified ifb_lo, ifb_high and ifb_signed? Or are you saying the IL spec document will be changed to indicate ffb only?

Right now, in order to do the equivalent to clz (ifb_high), I have to use the sequence 'bitreverse', 'ffb'. This just seems rather redundant obviously when the GPU has the ability to do this in one instruction.

Malcolm

[MicahVillmow]

malcolm3141,
This will be fixed in our upcoming release, the correct opcodes are ffb_hi, ffb_lo, ffb_shi.

[malcolm3141]

Fantastic! I can confirm those opcodes do work with Catalyst 10.7.

Thanks Michah!

Malcolm