Starting on Json file generation

hpvm/projects/keras/frontend/approxhpvm_translator.py

@@ -5,6 +5,7 @@ from frontend.promise_translator import PromiseRtTranslator
 from frontend.hpvm_dfg_translator import HPVMTranslator
 from frontend.weight_utils import dumpLabels, dumpData, dumpConvWeights, dumpFcWeights, dumpFcBias
 from frontend.utils import *
+from frontend.knobs import *
 import keras
 import os
 
@@ -203,11 +204,15 @@ class TensorRtTranslator:
     self.weight_str = ""
     self.program_str = ""
     self.input_str = ""
-    self.json_str = ""      # Used for Json gen
-    self.cur_height = 32    # Used for Json gen
-    self.cur_width = 32     # Used for Json egen
-    self.op_count = 1       # Used for Json gen
     self.filter_names = {}
+
+    # Used for Json gen
+    self.json_str = ""
+    self.knobs_str = ""
+    self.cur_height = 32    
+    self.cur_width = 32     
+    self.op_count = 1       
+    
     
 
 
@@ -215,7 +220,7 @@ class TensorRtTranslator:
 
     self.cur_height = data.shape[2]
     self.cur_width = data.shape[3]
-    DEBUG ("cur_height = ", self.cur_height, "  cur_width = ", self.cur_width, "\n")
+    DEBUG ("cur_height = ", self.cur_height, "  cur_width = ", self.cur_width, ", \n")
 
     
   def addConvOverheads(self, weights, padding, strides):
@@ -228,7 +233,8 @@ class TensorRtTranslator:
     flops = H_d * W_d * K_d
     DEBUG ("conv_flops =  ", flops)
 
-    self.json_str += "op" + str(self.op_count) + " : " + str(flops) + "\n"
+    self.json_str += "convolution_" + str(self.op_count) + " : " + str(flops) + ", \n"
+    self.knobs_str += "convolution_" + str(self.op_count) + " : ["  + conv_knobs + "], \n"
     self.op_count += 1
     
     self.cur_height = self.cur_height / strides[0]
@@ -242,7 +248,8 @@ class TensorRtTranslator:
     flops = weights.shape[0] * weights.shape[1]
     DEBUG ("dense_flops =  ", flops)
 
-    self.json_str += "op" + str(self.op_count) + " : " + str(flops) + "\n"
+    self.json_str += "linear_" + str(self.op_count) + " : " + str(flops) + "\n"
+    self.knobs_str += "linear_" + str(self.op_count) + " : ["  + baseline_knobs + "], \n"
     self.op_count += 1
         
     self.cur_height = 1
@@ -258,6 +265,14 @@ class TensorRtTranslator:
     
     DEBUG ("cur_height = ", self.cur_height, "  cur_width = ", self.cur_width, "\n")
 
+
+  def addBaselineKnob(self, op_name):
+
+    self.json_str += op_name + "_" + str(self.op_count) + " : 0, \n"
+    self.knobs_str += op_name + "_" + str(self.op_count) + " : ["  + baseline_knobs + "], \n"
+    self.op_count += 1
+
+    
     
     
   def getWeightStr(self):
@@ -451,7 +466,12 @@ class TensorRtTranslator:
       if layer_type == "Conv2D":
         self.addConvOverheads(weights, padding, strides)
   
-      
+      elif layer_type == "DepthwiseConv2D":
+        #self.json_str += "depthwise_convolution_" + str(self.op_count) + " : 0, \n"
+        #self.op_count += 1
+        self.addBaselineKnob("depthwise_convolution")
+
+    
 
     if layer_type == "Dense":
       input_var_name = self.getSingleInputName(cur_node)
@@ -481,13 +501,21 @@ class TensorRtTranslator:
       # NOTE: Changing output variable
       out_var_name1 = out_var_name2
 
+      #self.json_str += "add_" + str(self.op_count) + " : 0, \n"
+      # self.op_count += 1
+      self.addBaselineKnob("add")
+      
 
     if layer_type == "Activation":
       input_var_name = self.getSingleInputName(cur_node)
       
       inst_str = genActivationCallStr(input_var_name, out_var_name1, cur_node.activation_type)
       self.program_str += inst_str
-      
+
+      #self.json_str += cur_node.activation_type + "_" + str(self.op_count) + " : 0, \n"
+      #self.op_count += 1
+      self.addBaselineKnob(cur_node.activation_type)
+
     
     if self.hasActivation(cur_node) and layer_type != "Activation":
       activation_type = cur_node.activation_type
@@ -499,7 +527,11 @@ class TensorRtTranslator:
       if activation_type == "softmax":
         print ("Softmax canNOT be part of Dense/Conv Op. Insert: Activation('softmax');")
         sys.exit(0)
-        
+
+      #self.json_str += activation_type + "_" + str(self.op_count) + " : 0, \n"
+      #self.op_count += 1
+      self.addBaselineKnob(activation_type)
+      
 
     if layer_type == "BatchNormalization":
       input_var_name = self.getSingleInputName(cur_node)
@@ -514,6 +546,11 @@ class TensorRtTranslator:
       inst_str += "); \n"
       
       self.program_str += inst_str
+
+      #self.json_str += "batchnorm_" + str(self.op_count) + " : 0, \n"
+      #self.op_count += 1
+      self.addBaselineKnob("batchnorm")
+
       
       
     if layer_type == "Add":  
@@ -523,6 +560,10 @@ class TensorRtTranslator:
       inst_str += "tensorAdd(" + input_vars[0] + ", " + input_vars[1] + "); \n"
       self.program_str += inst_str
 
+      #self.json_str += "add_" + str(self.op_count) + " : 0, \n"
+      #self.op_count += 1
+      self.addBaselineKnob("add")
+
       
     if layer_type == "MaxPooling2D" or layer_type == "AveragePooling2D":  
       input_var_name = self.getSingleInputName(cur_node)
@@ -534,8 +575,16 @@ class TensorRtTranslator:
       pool_type = 0
       if layer_type == "MaxPooling2D":
         pool_type = "0"
+        #self.json_str += "maxpool_" + str(self.op_count) + " : 0, \n"
+        #self.op_count += 1
+        self.addBaselineKnob("maxpool")
+
       if layer_type == "AveragePooling2D":
-        pool_type = "1"     
+        pool_type = "1"
+        #self.json_str += "avgpool_" + str(self.op_count) + " : 0, \n"
+        #self.op_count += 1
+        self.addBaselineKnob("avgpool")
+
       
       # tensorPooling(input, pool_type, pool_h, pool_w, v_pad, h_pad, v_stride, h_stride)
       inst_str = "void* " + out_var_name1 + " = "
@@ -901,6 +950,7 @@ class TensorRtTranslator:
 
     f = open(dir_prefix + "/tuner.json", "w+")
     f.write(self.json_str)
+    f.write(self.knobs_str)
     f.close()
 
     

hpvm/projects/keras/frontend/knobs.py

+
+knobs_speedups = {}
+knobs_speedups[11] = 1
+knobs_speedups[12] = 1.5
+knobs_speedups[151] = 3
+knobs_speedups[152] = 3
+knobs_speedups[153] = 3
+knobs_speedups[154] = 3
+knobs_speedups[155] = 2.25
+knobs_speedups[156] = 2.25
+knobs_speedups[157] = 2.25
+knobs_speedups[158] = 2.25
+knobs_speedups[159] = 2.25
+knobs_speedups[160] = 2.25
+knobs_speedups[161] = 2
+knobs_speedups[162] = 2
+knobs_speedups[163] = 2
+knobs_speedups[164] = 2
+knobs_speedups[165] = 2
+knobs_speedups[166] = 2
+knobs_speedups[167] = 2
+knobs_speedups[168] = 2
+knobs_speedups[261] = 3
+knobs_speedups[262] = 3
+knobs_speedups[263] = 2.25
+knobs_speedups[264] = 2.25
+knobs_speedups[265] = 2.25
+knobs_speedups[266] = 2
+knobs_speedups[267] = 2
+knobs_speedups[268] = 2
+knobs_speedups[269] = 2
+
+
+conv_knobs = "12, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 261, 262, 263, 264, 265, 266, 267, 268, 269"
+
+baseline_knobs = "12"
+
+