OpenCL compiler generating mad instruction

krishnaraj — Mon, 13 Aug 2012 10:27:24 GMT

I have a compute instensive kernel performing lots of multiply and adds.
Am using opencl mad() function and float16 for the compiler to generate avx mad instruction.
But when I see the ASM dump from intel offline compiler, it shows mul and add instructions on YMM registers but no mad at all.

dump:
    ..................
    vmulps   YMM4, YMM3, YMMWORD PTR [R15 + R10 + 32864]
   vaddps   YMM2, YMM4, YMM2
   vunpckhps   YMM4, YMM0, YMM0
   vpermilps   YMM4, YMM4, 0
   vperm2f128   YMM4, YMM4, YMM0, 0
   vmulps   YMM5, YMM4, YMMWORD PTR [R15 + R10 + 32928]
   vaddps   YMM2, YMM5, YMM2
   vshufps   YMM5, YMM0, YMM0, 3
    .................

even tried cl-mad-enable (which was default) while building , but no change.
Am I missing something here?!

OpenCL compiler generating mad instruction

Alexander_Heinecke — Mon, 13 Aug 2012 13:45:02 GMT

Hi,

current AVX implementation doesn't have a mad instruction. This guy will be introduced with AVX2.

Alex

OpenCL compiler generating mad instruction

krishnaraj — Mon, 27 Aug 2012 11:44:34 GMT

Thanks for that. I saw FMA intrinsics on some intel page and didn't notice that it belonged to AVX2.
Hope its coming soon.

topic OpenCL compiler generating mad instruction in OpenCL* for CPU

OpenCL compiler generating mad instruction

OpenCL compiler generating mad instruction

OpenCL compiler generating mad instruction