[FEATURE] opRoofline in pytorch namespace

Applying roofline analysis for OP.

We aggregate all one op's kernel information for specific
op.

Signed-off-by: YushuoEdge <[email protected]>

src/amanda/profiler/metrics.py

@@ -1,10 +1,11 @@
+from itertools import count
 import numpy as np
 import pandas as pd
 import matplotlib.pyplot as plt
 from matplotlib.patches import Polygon
 from utils import findTopK
 
-def drawRoofline(hardwareTFlops, hardwareIntensity, X, Y):
+def drawRoofline(hardwareTFlops, hardwareIntensity, X, Y, op=False):
 	maxright=0
 	for x in X:
 		if x > maxright:
@@ -47,7 +48,10 @@ def func2(x):
 	for i in range(len(X)):
 		ax.scatter(X[i], Y[i], color="black", alpha=1, zorder=2, s=10)
 
-	plt.savefig("./Experiments/kernelRoofline_result.png")
+	if op == True:
+		plt.savefig("./Experiments/opRoofline_result.png")
+	else:
+		plt.savefig("./Experiments/kernelRoofline_result.png")
 
 
 
@@ -175,9 +179,12 @@ def kernelRoofline(supplyInfo, countData):
 		spOp = spAdd + spMul + spFma * 2
 		cyclesElapsed = dataList[index+5].gpuValue
 
+		# n*1e-6M / 1350MHz s
 		time = cyclesElapsed / (supplyInfo[0] * 1024 * 1024)
-		tflops = (spOp / time) / (1024 * 1024 * 1024)
+		# (n / ts) * 1e-12
+		tflops = (spOp / time) / (1024 * 1024 * 1024 * 1024)
 		kernelY.append(round(tflops, 2))
+		# n / mB
 		intensity = spOp / (dramRead + dramWrite)
 		kernelX.append(intensity)
 
@@ -314,4 +321,57 @@ def opInfoCounter(dataList, flopCount=True):
 	print(resCounter)
 	resCounter.to_csv("./Experiments/opInfoCounter_result.csv", index=False, sep=',')
 
-	return infoList	
+	return infoList
+
+# Op Roofline Analysis:
+# supplyInfo: [Frequency(MHz), Peak-Performance(TFlops), Throughput(GB/s)]
+def opRoofline(supplyInfo, countData):
+
+	# hardware TFLOPS & Intersity
+	hardwareTFlops = supplyInfo[1]
+	hardwareIntensity = (hardwareTFlops * 1000) / supplyInfo[2]
+
+	# Filter Count Data, aggregate the information
+	opPosition = []
+	opKernelNum = []
+	for i in range(len(countData)):
+		if countData[i].rangeName == "NEW OP":
+			opPosition.append(i)
+
+	for i in range(len(opPosition) - 1):
+		kernelNum = (opPosition[i+1] - opPosition[i] -1) / 6
+		opKernelNum.append(kernelNum)
+	kernelNum = (len(countData) - opPosition[-1] - 1) / 6
+	opKernelNum.append(kernelNum)
+
+	opX = []
+	opY = []
+	numValue = 6
+	for i in range(len(opPosition)):
+		if opKernelNum[i] == 0:
+			continue
+
+		totalDramRead = 0
+		totalDramWrite = 0
+		totalCyclesElapsed = 0
+		totalFlopCount = 0
+		for j in range(int(opKernelNum[i])):
+			kernelPosition = j * numValue + opPosition[i] + 1
+			totalDramRead += countData[kernelPosition].gpuValue
+			totalDramWrite += countData[kernelPosition+1].gpuValue
+			totalCyclesElapsed += countData[kernelPosition+5].gpuValue		
+			spAdd = countData[kernelPosition+2].gpuValue
+			spFma = countData[kernelPosition+3].gpuValue
+			spMul = countData[kernelPosition+4].gpuValue
+			totalFlopCount += spAdd + spMul + spFma * 2
+
+		# nM / 1350MHz s
+		time = totalCyclesElapsed / (supplyInfo[0] * 1024 * 1024)
+		# (nM / ts) * 1e-6 T  
+		tflops = (totalFlopCount / time) / (1024 * 1024 * 1024 * 1024)
+		opY.append(round(tflops, 2))
+		# nM / mMB
+		intensity = totalFlopCount / (totalDramRead + totalDramWrite)
+		opX.append(intensity)
+
+	drawRoofline(hardwareTFlops, hardwareIntensity, opX, opY)	

src/amanda/profiler/pytorch/__init__.py

@@ -1,2 +1,3 @@
 from amanda_tracer import *
-from amanda_counter import *
+from amanda_counter import *
+from profiler import *

src/amanda/profiler/pytorch/amanda_counter.py

@@ -2,7 +2,7 @@
 
 import amanda
 
-sys.path.append('../')
+sys.path.append('../..')
 from counter import counter
 
 class amandaCounter(amanda.Tool):

src/amanda/profiler/pytorch/amanda_tracer.py

@@ -2,7 +2,7 @@
 
 import amanda
 
-sys.path.append('../')
+sys.path.append('../..')
 from tracer import tracer
 
 class amandaTracer(amanda.Tool):

src/amanda/profiler/pytorch/profiler.py

@@ -4,7 +4,7 @@
 from amanda_counter import amandaCounter
 
 from utils import setConfigsMetric
-from metrics import kernelRoofline
+from metrics import kernelRoofline, opRoofline
 from torchMetrics import kernelInfo, opInfo
 
 class Profiler():
@@ -69,6 +69,12 @@ def showResults(self):
 			opInfo(self.opList, self.startTimeList, self.endTimeList, self.traceDataApi, self.traceDataRt, self.countData)
 			return
 
+		if self.__metric == "OpRoofline":
+			self.countData = self.counter.getCountData()
+			assert len(self.supplyInfo) == 3, "Please provide correct hardware parameters"
+			opRoofline(self.supplyInfo, self.countData)
+			return
+
 		sys.exit("Profiler.Metric: " + self.__metric + " not supported")
 
 

src/amanda/profiler/pytorch/counter_example.py → src/amanda/profiler/pytorch/test/counter_example.py

@@ -1,8 +1,10 @@
 from asyncore import file_dispatcher
 import amanda
 import torch
+import sys
 import torchvision
 
+sys.path.append("..")
 from amanda_counter import amandaCounter
 
 def main():

src/amanda/profiler/pytorch/kernelInfo_test.py → src/amanda/profiler/pytorch/test/kernelInfo_test.py

@@ -1,6 +1,10 @@
+import sys
+
 import amanda
 import torch
 import torchvision
+
+sys.path.append("..")
 from profiler import Profiler
 
 def main():

src/amanda/profiler/pytorch/kernelRoofline_test.py → src/amanda/profiler/pytorch/test/kernelRoofline_test.py

@@ -1,14 +1,17 @@
+import sys
+
 import amanda
 import torch
 import torchvision
-from pytorch.profiler import Profiler
+
+sys.path.append("..")
+from profiler import Profiler
 
 def main():
 
 	device = "cuda"
-
 	model = torchvision.models.resnet50().to(device)
-	x = torch.rand((32, 3, 227, 227)).to(device)
+	x = torch.rand((64, 3, 227, 227)).to(device)
 
 	metric = "KernelRoofline"
 	# Nvidia Geforce RTX 2080 Ti: 1350MHz, 13.45 Single-Precision TFlops, 616GB/s

src/amanda/profiler/pytorch/opInfo_test.py → src/amanda/profiler/pytorch/test/opInfo_test.py

@@ -1,6 +1,10 @@
+import sys
+
 import amanda
 import torch
 import torchvision
+
+sys.path.append("..")
 from profiler import Profiler
 
 def main():

src/amanda/profiler/pytorch/test/opRoofline_test.py

@@ -0,0 +1,29 @@
+import sys
+
+import amanda
+import torch
+import torchvision
+
+sys.path.append("..")
+from profiler import Profiler
+
+def main():
+
+	device = "cuda"
+
+	model = torchvision.models.resnet50().to(device)
+	x = torch.rand((32, 3, 227, 227)).to(device)
+
+	metric = "OpRoofline"
+	# Nvidia Geforce RTX 2080 Ti: 1350MHz, 13.45 Single-Precision TFlops, 616GB/s
+	supplyInfo = [1350, 13.45, 616]
+	profiler = Profiler(metric)
+	profiler.setConfigs(metric=metric, supplyInfo=supplyInfo)
+
+	with amanda.tool.apply(profiler.counter):
+		y = model(x)
+
+	profiler.showResults()
+
+if __name__ == "__main__":
+	main()

src/amanda/profiler/pytorch/tracer_example.py → src/amanda/profiler/pytorch/test/tracer_example.py

@@ -1,8 +1,10 @@
 from sklearn.preprocessing import KernelCenterer
 import amanda
 import torch
+import sys
 import torchvision
 
+sys.path.append("..")
 from amanda_tracer import amandaTracer
 
 def main():

src/amanda/profiler/tensorflow/__init__.py

@@ -1 +1,3 @@
-from amanda_tracer import *
+from amanda_tracer import *
+from amanda_counter import *
+from profiler import *

src/amanda/profiler/tensorflow/amanda_counter.py

@@ -4,7 +4,7 @@
 import tensorflow as tf
 
 import amanda
-sys.path.append('../')
+sys.path.append('../..')
 from counter import counter
 
 

src/amanda/profiler/tensorflow/amanda_tracer.py

@@ -4,7 +4,7 @@
 import tensorflow as tf
 
 import amanda
-sys.path.append('../')
+sys.path.append('../..')
 from tracer import tracer
 
 

src/amanda/profiler/tensorflow/counter_example.py → src/amanda/profiler/tensorflow/test/counter_example.py

@@ -1,10 +1,12 @@
 import os
+import sys
 import amanda
 import tensorflow as tf
 from tensorflow.python.client import timeline
+from examples.common.tensorflow.model.resnet_50 import ResNet50
 
+sys.path.append("..")
 from amanda_counter import amandaCounter
-from examples.common.tensorflow.model.resnet_50 import ResNet50
 
 def main():
 

src/amanda/profiler/tensorflow/kernelInfo_test.py → src/amanda/profiler/tensorflow/test/kernelInfo_test.py

@@ -1,8 +1,11 @@
+import sys
+
 import amanda
 import tensorflow as tf
+from examples.common.tensorflow.model.resnet_50 import ResNet50
 
+sys.path.append("..")
 from profiler import Profiler
-from examples.common.tensorflow.model.resnet_50 import ResNet50
 
 def main():
 

src/amanda/profiler/tensorflow/opInfo_test.py → src/amanda/profiler/tensorflow/test/opInfo_test.py

@@ -1,8 +1,11 @@
+import sys
+
 import amanda
 import tensorflow as tf
+from examples.common.tensorflow.model.resnet_50 import ResNet50
 
+sys.path.append("..")
 from profiler import Profiler
-from examples.common.tensorflow.model.resnet_50 import ResNet50
 
 def main():
 

src/amanda/profiler/tensorflow/tracer_example.py → src/amanda/profiler/tensorflow/test/tracer_example.py

@@ -1,10 +1,13 @@
+import sys
+
 from torch import int32
 import amanda
 import tensorflow as tf
 from tensorflow.python.client import timeline
+from examples.common.tensorflow.model.resnet_50 import ResNet50
 
+sys.path.append("..")
 from amanda_tracer import amandaTracer
-from examples.common.tensorflow.model.resnet_50 import ResNet50
 
 def main():
 

src/amanda/profiler/utils.py

@@ -67,5 +67,11 @@ def setConfigsMetric(metric, tracer, counter, flopCount=True):
 			counter.setKindFlag(0x5 | 0x42 << 42)
 		counter.setFilePath("./Experiments/kernel_metrics.txt")
 		return
+
+	# set configs for op roofline analysis
+	if metric == "OpRoofline" and flopCount:
+		counter.setKindFlag(0x1C05 | 0x1 << 42)
+		counter.setFilePath("./Experiments/kernel_metrics.txt")
+		return
 
 	sys.exit("Invalid Metric")