def get_reward(self, new_code, current_filename,
unroll_idx): #VF_idx,IF_idx):
'''Calculates the RL agent's reward. The reward is the
execution time improvement after injecting the pragma
normalized to -O3.'''
f = open(current_filename, 'w')
f.write(''.join(new_code))
f.close()
if self.compile:
if self.runtimes[current_filename][
self.current_pragma_idx][unroll_idx][0]: #VF_idx][IF_idx]:
runtime = self.runtimes[current_filename][
self.current_pragma_idx][unroll_idx][0] #VF_idx][IF_idx]
else:
runtime = get_runtime(self.new_rundir, new_code,
current_filename)
self.runtimes[current_filename][self.current_pragma_idx][
unroll_idx][0] = runtime #VF_idx][IF_idx]=runtime
if self.O3_runtimes[current_filename] == None:
reward = 0
logger.warning(
'Program ' + current_filename +
' does not compile in two seconds.' +
' Consider removing it or increasing the timeout parameter'
+ ' in utility.py.')
elif runtime == None:
#penalizing for long compilation time for bad VF/IF
reward = -9
else:
reward = (self.O3_runtimes[current_filename] -
runtime) / self.O3_runtimes[current_filename]
# In inference mode and finished inserting pragmas to this file.
if self.inference_mode and self.current_pragma_idx + 1 == self.num_loops[
current_filename]:
improvement = self.O3_runtimes[current_filename] / runtime
self.improvements.append(improvement)
geomean = 1
for imp in self.improvements:
geomean = geomean * (imp**(1 / len(self.improvements)))
print('benchmark: ', current_filename, 'O3 runtime: ',
self.O3_runtimes[current_filename], 'RL runtime: ',
runtime, 'improvement:',
str(round(improvement, 2)) + 'X',
'improvement geomean so far:',
str(round(geomean, 2)) + 'X')
#VF = self.vec_action_meaning[VF_idx]
#IF = self.interleave_action_meaning[IF_idx]
unroll = self.unroll_action_meaning[unroll_idx]
opt_runtime_sofar = self.get_opt_runtime(current_filename,
self.current_pragma_idx)
logger.info(
current_filename + ' runtime ' + str(runtime) + ' O3 ' +
str(self.O3_runtimes[current_filename]) + ' reward ' +
str(reward) + ' opt ' +
str(opt_runtime_sofar)) #+" VF "+str(VF)+" IF "+str(IF))
else:
# can't calculate the reward without compile/runtime.
reward = 0
return reward
def get_reward(self, new_code, current_filename):
f = open(current_filename, 'w')
f.write(''.join(new_code))
f.close()
if self.compile:
runtime = get_runtime(self.new_rundir, new_code, current_filename)
reward = (self.O3_runtimes[current_filename] -
runtime) / self.O3_runtimes[current_filename]
if self.inference_mode and self.current_pragma_idx + 1 == self.num_loops[
current_filename]: # in inference mode and finished inserting pragmas to this file
print('benchmark: ', current_filename, 'O3 runtime: ',
self.O3_runtimes[current_filename], 'RL runtime: ',
runtime)
else:
reward = 0 # can't calculate the reward without runtime
return reward