Newton-Raphson sqrt not accurate enough

[Coldfirex]

I got the above when running the following:
stress-ng --cpu 64 --cpu-method all --verify -t 5m --metrics-brief

What does this mean exactly?

[ColinIanKing]

It means that the computation did not reach the target sqrt value in the expected number of iterations to compute this. Which compiler and processor is being used?

[Coldfirex]

Host CPU is Intel i3-10105 running Proxmox
This was from a VM, Haiku, running g++ (2023_08_10) 13.2.0.

It might be because of the 64 setting? I have only allocated 2 vCPUs to this VM.

[ColinIanKing]

Please appy the following patch to the stress-ng source and re-run and let me know the value that is being printed out. For all cpus I've tested with gcc 13.2 it works, so I'm interested to see what values you get.

diff --git a/stress-cpu.c b/stress-cpu.c
index 2cc98325..2dc38712 100644
--- a/stress-cpu.c
+++ b/stress-cpu.c
@@ -477,7 +477,7 @@ static void OPTIMIZE3 TARGET_CLONES stress_cpu_nsqrt(const char *name)
 {
        int i;
        const long double precision = 1.0e-12L;
-       const int max_iter = 56;
+       const int max_iter = 128;
 
        for (i = 16300; i < 16384; i++) {
                const long double n = (long double)i;
@@ -495,6 +495,8 @@ static void OPTIMIZE3 TARGET_CLONES stress_cpu_nsqrt(const char *name)
                }
                rt = (lo + hi) / 2.0L;
 
+pr_inf("%d\n", j);
+
                if (g_opt_flags & OPT_FLAGS_VERIFY) {
                        const long double r2 = shim_rintl(rt * rt);

[ColinIanKing]

..it may be that the context switching in the kernel is causing register clobbering and messing up the result. Does this work with one instance of the stressor rather than 64?

[Coldfirex]

Just noticed our version is on 14.x. Let me see if we can package up a current version before we go any further here (as long as that makes sense to you).

[Coldfirex]

After upgrading to 0.16.05 its gone, so nevermind. :)