Layer fusion analysis running into errors
ssatyendras opened this issue · 2 comments
ssatyendras commented
Hi,
There seems to be some new issues after the recent updates to stream.
Trying to map resnet18 onnx model with hint_loops other than 1 seems to result in errors.
Error messages:
"Tensor Tensor(13, 13, 'O') was not found in any of the instances."
"Tensor Tensor(6, 2, 'O') was not found in any of the instances."
"There are no candidates to schedule."
Test scenario:
############################## Provide inputs ######################################
accelerator = "stream.inputs.examples.hardware.TPU_like_quad_core"
workload_path = "/stream/inputs/examples/workload/resnet18.onnx"
mapping_path = "stream.inputs.examples.mapping.tpu_like_quad_core"
############################## Define variables for run ############################
CN_define_mode = 1 # manually define outer-CN loops
hint_loops = [("OY", 2)] # [("OY", 'all')] [("OY", 1)] [("OY", 2)] [("OY", 3)]
nb_ga_individuals = 100 # number of individuals in each generation
nb_ga_generations = 10 # number of genetic algorithm generations
OBJECTIVE = "latency" ## latency memory
mainstage = MainStage(
[ # Initializes the MainStage as entry point
AcceleratorParserStage, # Parses the accelerator
StreamONNXModelParserStage, # Parses the ONNX Model into the workload
# UserDefinedModelParserStage, # Parses the user-defined Model into the workload
GenerateCNWorkloadHybridStage,
IntraCoreMappingStage,
InterCoreMappingStage,
],
accelerator=accelerator, # required by AcceleratorParserStage
workload_path=workload_path, # required by ModelParserStage
mapping_path=mapping_path, # required by ModelParserStage
loma_lpf_limit=6, # required by LomaStage
nb_ga_individuals=nb_ga_individuals, # number of individuals in each genetic algorithm generation
nb_ga_generations=nb_ga_generations, # number of genetic algorithm generations
node_hw_performances_path=node_hw_performances_path, # saved node_hw_performances to skip re-computation
plot_hof=True, # Save schedule and memory usage plot of each individual in the Genetic Algorithm hall of fame
plot_file_name=plot_file_name,
plot_full_schedule=plot_full_schedule,
plot_data_transfer=plot_data_transfer,
cn_define_mode=CN_define_mode,
hint_loops=hint_loops,
scheduler_candidate_selection=OBJECTIVE,
operands_to_prefetch=[],
)
Kindly take a look.
Thanks & regards,
Siva
Andful commented
I suspect the error is related to #7. The pull request is still not merged. You could try using the branch of that pull request.
asyms commented
Hi @ssatyendras,
Has this issue been resolved? The PR has been merged in since.