All geek questions in one place

ARM NEON: What pairs of instructions should wait for a write?

The ARM NEON documentation says:

[...] some pairs of commands may need to wait until the value is written back to the register file.

I have not come across a list that defines pairs of instructions that can use redirected results and pairs of commands that have to wait for a record.

Does anyone know the table or documentation that lists these pairs?

+4
source share
3 answers

Broadly speaking, what you reasonably expect is forward, forward. vmul.f32 goes to vadd.f32, etc.

I do not believe that the exact shipping routes are accurately documented anywhere you look. In any case, I did not find them. If you find them, be sure to tell us where. Of course, it is not difficult to determine for any pair of instructions whether the forwarding occurs, but this is not a general solution. Unfortunately.

+1
source

Does anyone know the table or documentation that lists these pairs?

These pairs are over 9000 and all of them cannot be listed.
For instance:

VADD.F32 q0,q0,q1 VMUL.F32 q3,q0,q2

the first command writes the result in the 4th cycle, while the second command requires (q0) as the source in the 2nd cycle, since the source is not ready yet, there is a stall (or conveyor hole) between these two instructions.

To calculate these kiosks, you can use the following online tool:
http://pulsar.webshaker.net/ccc/result.php?lng=us

+1
source

The integer is multiplied.

The section at the end of http://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.ddi0344k/ch16s06s03.html is useful:

If multiply-accumulate should be multiplied or another multiply-accumulate and depends on the result of the first instruction, then if the relationship between both instructions is the same type and size, the processor uses a special battery forwarding multiplier. This special transfer means the multiplication of instructions can be issued alternately, because the result of the first instruction in N5 is sent to the battery of the second instruction in N4. If the size and type of instructions are not specified, then N3 requires Dd or Qd. This applies to combinations of the VMLA, VMLS, VQDMLA and VQDMLS multiply-accumulate commands, and the VMUL and VQDMUL multiply commands

Do not assume that floating point multiplication accumulates work in the same way. I have not used NEON floating point instructions for any critical performance, so I cannot offer you any experience, but make sure you read and understand the note at the end of http://infocenter.arm.com/help/index.jsp ? topic = / com.arm.doc.ddi0344k / BCGDCECC.html

+1
source

Source: https://habr.com/ru/post/1384432/

More articles:


All Articles