[igt-dev] [PATCH i-g-t] benchmarks/gem_blt: fix baseline estimation

[Kamil Konieczny]

Hi Mauro,

small nitpicks on commit message.

On 2022-05-19 at 09:41:10 +0200, Mauro Carvalho Chehab wrote:
> From: Mauro Carvalho Chehab <mchehab at kernel.org>
> 
> This test is expected to run under a certain time defined by a command
> line parameter (by default, 2s).
> 
> In order to achive the specified time, the baseline() function tries to
> estimate the value of a counter that will get the minimal amount of
> interaction to wait for > 0.1s.
> 
> However, currently, the baseline estimation is broken, returning a too high
------------------------------------------------------------------------ ^
s/high/big/

> number, as it is measuring the memcpy() time, instead of actually measuring
> each loop's interaction.

/each loop's interaction./gem_execbuf()./

> 
> Due to that, a default test without passing any command line parameter
> would take more than 1.5 years to output the first benchmark data!
> 
> Fix it by using the same logic that it is inside run() at the baseline
> time estimation.
> 
> On ADL-P, before this patch, baseline would return 399242693. After

s/On ADL-P, before/Before/
s/would/could/

> it, it now return 20 - meaning an interval of about 120 ms, which

Well, both numbers means the same interval _if_ used for the
same syscall.

With that fixed you can add my r-b tag.

--
Kamil

> should be good enough to ensure that the tests won't take too long.
> 
> Signed-off-by: Mauro Carvalho Chehab <mchehab at kernel.org>
> ---
>  benchmarks/gem_blt.c | 18 +++++-------------
>  1 file changed, 5 insertions(+), 13 deletions(-)
> 
> diff --git a/benchmarks/gem_blt.c b/benchmarks/gem_blt.c
> index 424ce8e75913..bd8264b4e896 100644
> --- a/benchmarks/gem_blt.c
> +++ b/benchmarks/gem_blt.c
> @@ -57,29 +57,21 @@ elapsed(const struct timespec *start, const struct timespec *end)
>  	return (end->tv_sec - start->tv_sec) + 1e-9*(end->tv_nsec - start->tv_nsec);
>  }
>  
> -static int baseline(uint64_t bytes, int milliseconds)
> +static int baseline(int fd, struct drm_i915_gem_execbuffer2 execbuf, int milliseconds)
>  {
>  	struct timespec start, end;
> -	const int size = 64*1024*1024;
>  	int count = 0;
> -	void *mem;
> -
> -	mem = malloc(size);
> -	if (mem == NULL)
> -		return 1;
>  
>  	clock_gettime(CLOCK_MONOTONIC, &start);
>  	do {
> -		memset(mem, count, size);
> +		gem_execbuf(fd, &execbuf);
>  		count++;
>  		clock_gettime(CLOCK_MONOTONIC, &end);
> -		if (elapsed(&start, &end) > 0.1)
> +		if (elapsed(&start, &end) > (milliseconds / 1000.))
>  			break;
>  	} while (1);
>  
> -	free(mem);
> -
> -	return ceil(1e-3*milliseconds/elapsed(&start, &end) * count * size / bytes);
> +	return count;
>  }
>  
>  static int gem_linear_blt(int fd,
> @@ -260,7 +252,7 @@ static int run(int object, int batch, int time, int reps, int ncpus, unsigned fl
>  		execbuf.flags |= I915_EXEC_NO_RELOC;
>  
>  	/* Guess how many loops we need for 0.1s */
> -	count = baseline((uint64_t)object * batch, 100) / ncpus;
> +	count = baseline(fd, execbuf, 100) / ncpus;
>  	if (flags & SYNC) {
>  		time *= count / 2;
>  		count = 1;
> -- 
> 2.36.1
>