`during` observer callbacks are silently skipped with the Compiler backend

[askorikov]

Since kernel_finished is always True for the Complier backend, the during() observer callbacks never get executed:

kernel_tuner/kernel_tuner/core.py

Lines 410 to 412 in 96a6abc

	while not self.dev.kernel_finished():
	for obs in self.benchmark_observers:
	obs.during()

This may not be obvious for inexperienced users, so maybe it's worth adding a warning about this behavior?

Reproduction example

import numpy
from kernel_tuner import tune_kernel
from kernel_tuner.observers.nvml import NVMLObserver
from collections import OrderedDict

kernel_string = """ 
#include <omp.h>

typedef float vfloat __attribute__ ((vector_size (vecsize*4)));

extern "C" float vector_add(vfloat *c, vfloat *a, vfloat *b, int n) {
    double start = omp_get_wtime();
    int chunk = n/(vecsize*nthreads);

    #pragma omp parallel num_threads(nthreads)
    {
        int offset = omp_get_thread_num()*chunk;
        for (int i = offset; i<offset+chunk && i<n; i++) {
            c[i] = a[i] + b[i];
        }
    }

    return (float)((omp_get_wtime() - start)*1e3);
}
"""

size = 72 * 1024 * 1024

a = numpy.random.randn(size).astype(numpy.float32)
b = numpy.random.randn(size).astype(numpy.float32)
c = numpy.zeros_like(b)
n = numpy.int32(size)

args = [c, a, b, n]

tune_params = OrderedDict()
tune_params["nthreads"] = [1, 2, 3, 4, 8, 12, 16, 24, 32]
tune_params["vecsize"] = [1, 2, 4, 8, 16]

answer = [a + b, None, None, None]


metrics = OrderedDict()
metrics['power'] = lambda p: p['nvml_power']
    
tune_kernel(
    "vector_add",
    kernel_string,
    size,
    args,
    tune_params,
    answer=answer,
    compiler_options=["-fopenmp", "-O3"],
    observers=[NVMLObserver(['nvml_power'])],
    metrics=['nvml_power'],
)

nthreads=1, vecsize=1, time=35.764, power=0
nthreads=1, vecsize=2, time=34.871, power=0
nthreads=1, vecsize=4, time=33.393, power=0
nthreads=1, vecsize=8, time=34.477, power=0
...