Check out #854. It should be even faster by requesting only one usercmd_t per frame, and doesn't require ABI changes.

I believe this issue might need an update since mentioned issue was closed?

We may still want to implement this even if we already merged a workaround that fixes the performance problem. What this can bring is to make the code simpler by calling a function once at the beginning of the frame processing, instead of calling a function in various places and in many times.

The good thing with the workaround is that it achieves the same performance without breaking engine compatibility, we may still want to refactor and clean-up the code anyway. The priority is just far more lower.

This one may be less needed since an alternative code already exists and was merged in #854:

• Keep a cache of usercmd_t's in cgame #854

I like the fact this implementation makes the code less convoluted on game side.

This PR makes only sense if sizeof( userCmds_t ) is smaller than the kernel buffer, because if the single call to fetch all usercmd_t per frame splits the packets, the caching code only fetching one usrcmd_t per frame would be faster.

For reference:

#define CMD_BACKUP 64

#define USERCMD_BUTTONS 16

struct usercmd_t
{
	int         serverTime;
	int         angles[ 3 ];

	signed char forwardmove, rightmove, upmove;
	dtType_t    doubleTap;

	byte        weapon;
	byte        flags;

	byte        buttons[ USERCMD_BUTTONS / 8 ];
};

using userCmds_t = std::array<usercmd_t, CMD_BACKUP>;

So, sizeof( userCmds_t ) is 1536.

Actually one advantage of this implementation is that by turning the array into a vector, one could easily make the amount of command backups fetched for prediction very easy to change by the user with a cvar.

I noticed this:

void CG_DrawActiveFrame( int serverTime, bool demoPlayback )
{
	// …

	cg.currentCmdNumber = trap_GetCurrentCmdNumber();

	// …

	cg.currentCmd = trap_BatchGetUserCmds( cg.userCmds );
	
	// …
}

I guess that at some point we will be able to merge those two trap calls into a single one.

But really, those trap calls now count for less than 1%, that's very good!

This Orbit screenshot was taken while running a branch featuring all my currently implement batch trap calls.

In fact cg.currentCmdNumber is not used anymore ! We can just drop everything related to it! 🤪️

This will even divide by two the time spent in trap calls in this function! 😃️

Yay!