diff --git a/llvm/projects/soc_simulator/src/driver_new_config_fp16_repl.py b/llvm/projects/soc_simulator/src/driver_new_config_fp16_repl.py
index f53573f7cde9420400194827d55d84d69e2ace5b..d2d175725397994c8e76dd33bf79d77fe652aa3e 100644
--- a/llvm/projects/soc_simulator/src/driver_new_config_fp16_repl.py
+++ b/llvm/projects/soc_simulator/src/driver_new_config_fp16_repl.py
@@ -189,7 +189,6 @@ class Driver:
curr_conf_results.append((layer_as_lst[1], layer_results))
line = config_file.readline().strip()
continue
-
layer_ind = int(layer_as_lst[0]) - 1
layer_table_data = self.__tensor_layers[layer_ind]
layer_name = layer_table_data["Name"]
@@ -233,7 +232,6 @@ class Driver:
elif precision_type == "samp":
approx_type = Driver.ApproxTypes.SAMP
curr_layer = Driver.PrecisionTypes.FP16
- print(curr_layer, prev_layer)
quant_time, quant_energy = self.__quantize(precision_type, op_number, curr_layer, prev_layer, tensor_count, layer_table_data)
if quant_time != 0:
assert i == 2 #and layer_ind == 0
@@ -256,9 +254,8 @@ class Driver:
has_quantized = False
for layer_ind, (hardware, layer) in enumerate(curr_conf_results):
- if len(layer) == 1 and layer[0][2].find("softmax") != -1: continue
+ if layer[0][2].find("softmax") != -1: continue
fp16_layer = []
- #print(layer_ind, hardware, layer)
layer_table_data = self.__tensor_layers[layer_ind]
layer_name = layer_table_data["Name"]
@@ -302,7 +299,6 @@ class Driver:
else:
lookup_key = "_" + precision_type + str(op_number) + "_"
- print(curr_layer)
if curr_layer == Driver.PrecisionTypes.FP32:
time_key = "h2f%stime" % lookup_key
energy_key = "h2f%senergy" % lookup_key
@@ -311,8 +307,6 @@ class Driver:
energy_key = "f2h%senergy" % lookup_key
time = tensor_op_row[time_key]
energy = tensor_op_row[energy_key]
- print(time_key, energy_key)
- print("Quantization: (%f, %f)" % (time, energy))
return (time, energy)
@@ -330,7 +324,7 @@ class Driver:
elif Driver.is_fc(layer_name):
rows_a = layer_data["RA"]
cols_a = layer_data["CA"]
- rows_b = cols_
+ rows_b = layer_data["RB"]
cols_b = layer_data["CB"]
else:
print("PROMISE can't run whatever this layer is.")
@@ -349,9 +343,6 @@ class Driver:
def __run_gpu_simulation(self, curr_layer, layer_name, tensor_ind, \
approx_type = None, knob_number = None):
tensor_info = self.__tensor_table[layer_name][tensor_ind]
- #print(tensor_info)
- #print(layer_name)
- #print(tensor_ind)
time_key = None
energy_key = None
@@ -427,7 +418,7 @@ class Driver:
def get_final_times_energies_conf(curr_conf, curr_conf_name):
final_time = final_energy = 0
-
+
final_conf = [] # List (conf) of lists (layers) of tuples (operation data)
#for hardware, layer in self.fp16_baseline:
@@ -440,24 +431,36 @@ class Driver:
final_conf_layer.append((None, None, tensor_op))
continue
# layer name, operation name, val name
- baseline_time = self.fp16_baseline[layer_ind][1][tensor_ind][0]
- baseline_energy = self.fp16_baseline[layer_ind][1][tensor_ind][1]
- baseline_op = self.fp16_baseline[layer_ind][1][tensor_ind][2]
- #print(baseline_time, baseline_energy, baseline_op)
- #print(op_time, tensor_op)
+ if tensor_op.find("promise") != -1: # compute sum of entire fp16 baseline layer
+ baseline_time = 0
+ baseline_energy = 0
+ baseline_op = []
+ baseline_layer = self.fp16_baseline[layer_ind][1]
+
+ for op_time, op_energy, tensor_op in baseline_layer:
+ baseline_time += op_time
+ baseline_energy += op_energy
+ baseline_op.append(tensor_op)
+ else: # look at the individual tensor operation as before
+ baseline_time = self.fp16_baseline[layer_ind][1][tensor_ind][0]
+ baseline_energy = self.fp16_baseline[layer_ind][1][tensor_ind][1]
+ baseline_op = self.fp16_baseline[layer_ind][1][tensor_ind][2]
+
final_tensor_op = tensor_op
- #print(op_time > baseline_time)
if op_time > baseline_time:
- #print("**************** BIGGER ******************")
- #print(curr_conf_name)
- #print(baseline_time, baseline_energy, baseline_op, layer_ind)
- #print(op_time, tensor_op, layer_ind)
+ print("**************** BIGGER ******************")
+ if tensor_op.find("promise") != -1:
+ print("WARNING: PROMISE")
+ print(curr_conf_name)
+ print(baseline_time, baseline_energy, baseline_op, layer_ind)
+ print(op_time, tensor_op, layer_ind)
final_time += baseline_time
final_energy += baseline_energy
final_tensor_op = baseline_op
else:
final_time += op_time
final_energy += op_energy
+
final_conf_layer.append((None, None, final_tensor_op)) # Don't care about the times and energies when writing
final_conf.append((hardware, final_conf_layer))
#print("\n")