what is the mechanism of SRA/SLA when using Fixed Point

[saber cheng]

Hi all,

Recently, i received a strange problem.

when using spatial and in the Accel{} region, for example, i define "val x = Reg[FixPt[TRUE, _16, _16]]" and i want to do SRA.

so i define another parameter "val shift = x.to[FixPt[TRUE, _16, _0]]", i use "result" to receive the return value of "x >> shift".

But here is a strange problem, if i set x to 0, then the result will be 0 rather tnan x. And if the x doesnt equal to 0, for example 1 or 2, the the result will be what it should be like x/2 or x/4.

So can anyone help me with this

[Matt F]

Hi Saber,

Looks like you found a bug.  I'll fix this shortly.

Just to clarify, you said that doing  x >> x.to[FixPt[TRUE,_16,_0]]  gives 0, but I assume you meant anything shifted by 0 gives 0.

When you shift by a non-constant value, it replaces that shift with a loop structure whose counter is set by the shift amount.  In general, if you can get away with creating a list of results each shifted by a different amount and then using a mux to select the correct result based on shift, this would probably give you better performance.

[saber cheng]

sorry , what i expressed was a little confusing

let me re-explain it

in the Accel{} region

i have a register "val x = Reg[FixPt[TRUE,_16,_16]]"

then i have const or input value "val shift = ArgIn[FixPt[TRUE, _16, _0]]"

so when the shift is set to zero, then the return value of "x >> shift" will always be 0

hope this time i can make it clear, 

here i will attach my code 

[Matt F]

Thanks.  This should be fixed now (pull if you are using spatial repo, or just recompile if you are using spatial-quickstart).  Let me know if you still have any problems.