Archives Discussions

[berathebrain]

I have tried to change the reduction kernel sample to work with float4 numbers. It turns out that the kernel produces some minor errors when adding numbers. The reference CPU reduction is different from OpenCL reduction. It differs by 0.01% but if I have like a million values then the error is pretty huge.

Can someone confirm this, or am I doing something wrong.

With float works but with float4 it doesn't;

Tried the CPU context and GPU.

I have Radeon 4870, Windows 7 64bit.

[berathebrain]

Actually, if the length of input data is somewhat large, it also fails for simple float. I have just changed the kernels and everything in Reduction.cpp from uint4 to float and sometimes I get good results and sometimes bad.

Does anyone know where is the problem?

[redditisgreat]

It's obviously rounding errors.  If the CPU accumulates results in the 80 bit floating point registers you get less rounding errorst than if you do it with 32 bit variables in the GPU or the SSE unit.

[genaganna]

Originally posted by: berathebrain Actually, if the length of input data is somewhat large, it also fails for simple float. I have just changed the kernels and everything in Reduction.cpp from uint4 to float and sometimes I get good results and sometimes bad.

Does anyone know where is the problem?

Berathebrain,

       Problem is in floating point arithmetic.  Associative property is not a valid property for floating point addition and multiplication.

      EX :

        1.  (  a + b ) + c  is not equal to  a + ( b + c )

        2.  (  a * b ) * c  is not equal to  a * ( b * c )