Archives Discussions

[rick_weber]

Is there a way to get the compiler to emit dmad instructions without calling mad() or fma()? I looked at the fma macro and it does the following:

It seems the 4 mov instructions are extraneous and the in0, in1, in2, and out0 registers can be directly fed into the dmad instruction.

mdef(358)_out(1)_in(3) mov r0, in0 mov r1, in1 mov r2, in2 dmad r0.xy__, r0.xy, r1.xy, r2.xy mov out0, r0 mend

[MicahVillmow]

rick.weber, this is something that will show up in the upcoming release. It did not make it into 2.2.

Also, the mov instructions in IL are not important if they do not show up in the ISA.

[rick_weber]

The movs get optimized away in backend compilation then?

[MicahVillmow]

Yes it should. However, if it is not the case, then it is a problem that we need to look into fixing.

[rick_weber]

The code I observed was generated by using clGetProgramInfo and storing the binary to a file. Is there a way to instead view the device-specific assembly code? I think most of the movs are getting optimized away as I'm getting over 50% of double precision peak in a DGEMM using fma().

[MicahVillmow]

set the environment variable GPU_DUMP_DEVICE_KERNEL=2 will dump the ISA for each kernel. Another solution is to use Stream Kernel Analyzer to get the ISA.

[rick_weber]

Thanks! So it appears those mov instructions indeed were optimized away in backend compilation.