Will SETI ever return?

Thanks for asking.

Twiddle_dee is a small table of constant values in the code that is used in FFT calculations (to bring out repeating patterns of signals).

On NVIDIA constant values placed in __constant memory area should be accessed simultaneously at one memory address by all GPU threads (2000-5000 of them).

Einstein software fetches the twiddle_dee values almost randomly. So a performance penalty occurs: The nearly instantaneous parallel fetch of the value is turned into a sequential one. Not one clock cycle, but thousands of clock cycles: all computing waits for the last read from memory to be able to continue.

The FIX is to define twiddle_dee (a small buffer of constants) to reside in global memory and thus not be fetched through constant cache sequentially. The read from global memory can benefit from caching and nearby access of the values that were fetched with some earlier read, A read of one address fills the cache with more values ahead. So : Global memory can be served to the threads from 'random' addresses much faster.

The 'slowdown bug' was not really a bug. It is quite common to think that constants should always be placed in constant memory space or at least be fetched through constant cache (a programming technique used with GPUs).

--

Petri

p.s. The access pattern has localIndex(0) (lI.x, aka thread id) as its base variable. So every thread reads from a different address.

...
"int xInd=lI.x+localExtent.x*(lI.y%wTE.y);"
"int yInd=lI.y/wTE.y;"
"int gInd=xInd+rowSizeinUnits*yInd;"
"int TW3Ind=(groupIndex.x*wTE.x+xInd)*(currDimIndex*wTE.y*wgUnroll+yInd);"
"int TW3add=(groupIndex.x*wTE.x+xInd)*wTE.y;"
"tileIn+=gInd;"
"\n#pragma unroll 16\n"
"for(int t=0;t<wgUnroll;t++)"
"{"
"tmp=SLf2(tileIn,sizeof(float2)*t*rowSizeinUnits*wTE.y);"
"int u=TW3Ind;"
"TW3Ind+=TW3add;"

"global float2 *restrict t0=&twiddle_dee[0][0],*restrict t1=&twiddle_dee[1][0],*restrict t2=&twiddle_dee[2][0];"
"float2 result=GLf2(&t0[u & 0xff]),T1=GLf2(&t1[(u>>8)&0xff]),T2=GLf2(&t2[(u>>16)&0xff]);"
"result=(float2)((result.x*T1.x-result.y*T1.y),"

...