Cycle count for each opcode

[Gravis]

I need to know how many cycles each opcode takes (sans penalties or delays) in order to make the multicore emulator accurate. from the pipeline description, i can figure out/guess most cycle counts but i really want 100% accuracy. so Adapteva guys... how about it?

[aolofsson]

All instructions follow the same path until E1:

fe: PC sent to instruction memory
im: instructions come back from memory
de: instructions decoded
ra: register values look up and operand prep (could be immediates as well)
e1: most instructions completes (!(ldr | fpu))
e2: load from local memory completes
e3: fpu (or secondary ialu) instructions complete when rounding mode in CONFIG is set to truncation
e4: fpu instructions complete when rounding mode is round to nearest even.

Andreas

[Gravis]

so GIE/GID/NOP/IDLE/BKPT/TRAP are all 5 cycles?

[timpart]

[ysapir]

@Tim,

What rounding do you expect with integer math? What exactly would an IMADD operation round (or truncate)?

It would be a valid question if we supported saturation, but the way it is, it's always truncation, or more exactly, wrap-around.

[timpart]

I'm not expecting any rounding with secondary integer, can't see how it could apply.

What I meant was, if the configuration is set to secondary integer with rounding do those instructions take until E4, or do they finish at E3? So if we are using the secondary integer instructions do we need to set configuration to truncate?

Tim

[aolofsson]

To make secondary int instructions finish at e3 you need to set the truncate bit in the Config register.

Andreas