[BUGFIX] Mapping mechanism

In this commit, two kinds of mapping problems are fixed: (1) OP-KERNEL Mapping. The mapping mechanisms has been rewritten here. The previous one would cause some errors. (2) Tracer-Counter in tensorflow. We found that when tracer and counter run on a computational graph in a process, they interact with each other. And the kernel is not exactly same when running some ops. So we define metrics xxInfoTracer and xxInfoCounter to measure separately. Signed-off-by: YushuoEdge <[email protected]>
uchuhimo · Apr 28, 2022 · 138124c · 138124c
1 parent 547cf32
commit 138124c
Show file tree

Hide file tree

Showing 17 changed files with 124 additions and 77 deletions.
diff --git a/src/amanda/profiler/metrics.py b/src/amanda/profiler/metrics.py
@@ -3,6 +3,7 @@
 import pandas as pd
 import matplotlib.pyplot as plt
 from matplotlib.patches import Polygon
+from torch import take
 from utils import findTopK
 
 def drawRoofline(hardwareTFlops, hardwareIntensity, X, Y, op=False):
@@ -67,28 +68,32 @@ def kernelInfoTracer(opList, timeList, apiList, rtList):
 		if x.kind == "KERNEL" or x.kind == "CONC KERNEL":
 			kernelList.append(x)
 
+	infoTime = []
+	for i in range(len(opList)):
+		infoTime.append(["OP", opList[i], timeList[i]])
+	for i in range(len(launchKernelApiList)):
+		infoTime.append(["KERNEL", launchKernelApiList[i].name, launchKernelApiList[i].startTime])
+
+	def takeTime(elem):
+		return elem[2]
+
+	infoTime.sort(key=takeTime)
+
 	# Tracer match kernel and op
 	infoList = []
-	opIndex = 0
+	opIndex = -1
 	kernelIndex = 0
-	timeListLen = len(timeList)
-	for i in range(len(launchKernelApiList)):
-		record = []
-		if (opIndex < timeListLen - 1 and launchKernelApiList[i].startTime > timeList[opIndex + 1]):
+	kernelNum = 0
+	for i in range(len(infoTime)):
+		if infoTime[i][0] == "OP":
 			opIndex += 1
 			kernelIndex = 0
-
-		record.append(opIndex)
-		record.append(opList[opIndex])
-
-		record.append(kernelIndex)
-		record.append(kernelList[i].kind)
-		record.append(kernelList[i].name)
-		record.append(launchKernelApiList[i].durationTime)
-		record.append(kernelList[i].durationTime)
-
-		infoList.append(record)
-		kernelIndex += 1
+		else:
+			infoList.append([opIndex, opList[opIndex], kernelIndex, kernelList[kernelNum].kind,
+							kernelList[kernelNum].name, launchKernelApiList[kernelNum].durationTime,
+							kernelList[kernelNum].durationTime])
+			kernelIndex += 1
+			kernelNum += 1
 
 	# Find Top-K kernel according to kernel execution time
 	k = min(20, len(kernelList))
@@ -206,33 +211,26 @@ def opInfoTracer(opList, startTimeList, endTimeList, apiList, rtList):
 			kernelList.append(x)
 
 	# Calculate number of kernels for each op
-	kernelNumList = []
-	kernelCount = 0
-	opIndex = 0
-	kernelIndex = 0
-	timeListLen = len(startTimeList)
-
-	while opIndex < timeListLen - 1 and startTimeList[opIndex] < launchKernelApiList[0].startTime and startTimeList[opIndex+1] < launchKernelApiList[0].startTime:
-		kernelNumList.append(0)
-		opIndex += 1
+	infoTime = []
+	for i in range(len(opList)):
+		infoTime.append(["OP", opList[i], startTimeList[i]])
+	for i in range(len(launchKernelApiList)):
+		infoTime.append(["KERNEL", launchKernelApiList[i].name, launchKernelApiList[i].startTime])
 
-	while kernelIndex < len(launchKernelApiList):
-		if (opIndex < timeListLen - 1 and launchKernelApiList[kernelIndex].startTime > startTimeList[opIndex + 1]):
-			kernelNumList.append(kernelCount)
-			opIndex += 1
-			kernelCount = 0
-			continue
-		kernelCount += 1
-		kernelIndex += 1
-	kernelNumList.append(kernelCount)
+	def takeTime(elem):
+		return elem[2]
+	infoTime.sort(key=takeTime)
 
-	while opIndex < timeListLen - 1:
-		kernelNumList.append(0)
-		opIndex += 1
-
-	for i in range(len(opList)):
-		print(opList[i])
-		print(kernelNumList[i])
+	kernelNumList = []
+	kernelCountOp = 0
+	for i in range(len(infoTime)):
+		if infoTime[i][0] == "OP":
+			kernelNumList.append(kernelCountOp)
+			kernelCountOp = 0
+		else:
+			kernelCountOp += 1
+	kernelNumList.append(kernelCountOp)
+	kernelNumList = kernelNumList[1:]
 
 	# Get information for each op
 	kernelCount = 0

diff --git a/src/amanda/profiler/pytorch/__init__.py b/src/amanda/profiler/pytorch/__init__.py
@@ -1,3 +1,3 @@
 from amanda_tracer import *
 from amanda_counter import *
-from profiler import *
+from amanda_profiler import *
diff --git a/src/amanda/profiler/pytorch/amanda_counter.py b/src/amanda/profiler/pytorch/amanda_counter.py
@@ -17,9 +17,6 @@ def forward_instrumentation(self, context: amanda.OpContext):
 		op = context.get_op()
 		self.opCount += 1
 
-		# if self.opCount > 10:
-		# 	return
-
 		self.opList.append(op.__name__)
 		context.insert_before_op(
 			self.start_profiling,

diff --git a/src/amanda/profiler/pytorch/profiler.py → ...manda/profiler/pytorch/amanda_profiler.py b/src/amanda/profiler/pytorch/profiler.py → ...manda/profiler/pytorch/amanda_profiler.py
@@ -7,7 +7,7 @@
 from metrics import kernelRoofline, opRoofline
 from torchMetrics import kernelInfo, opInfo
 
-class Profiler():
+class amandaProfiler():
 	def __init__(self, metric) -> None:
 		self.__metric = metric
 

diff --git a/src/amanda/profiler/pytorch/amanda_tracer.py b/src/amanda/profiler/pytorch/amanda_tracer.py
@@ -18,9 +18,6 @@ def forward_instrumentation(self, context: amanda.OpContext):
 		op = context.get_op()
 		self.opCount += 1
 
-		# if self.opCount > 10:
-		# 	return
-
 		self.opList.append(op.__name__)		
 		context.insert_before_op(
 			self.init_trace,

diff --git a/src/amanda/profiler/pytorch/test/kernelInfo_test.py b/src/amanda/profiler/pytorch/test/kernelInfo_test.py
@@ -5,7 +5,7 @@
 import torchvision
 
 sys.path.append("..")
-from profiler import Profiler
+from amanda_profiler import amandaProfiler
 
 def main():
 
@@ -15,7 +15,7 @@ def main():
 	x = torch.rand((32, 3, 227, 227)).to(device)
 
 	metric = "KernelInfo"
-	profiler = Profiler(metric)
+	profiler = amandaProfiler(metric)
 	profiler.setConfigs(metric=metric, supplyInfo=[])
 
 	with amanda.tool.apply(profiler.counter):

diff --git a/src/amanda/profiler/pytorch/test/kernelRoofline_test.py b/src/amanda/profiler/pytorch/test/kernelRoofline_test.py
@@ -5,7 +5,7 @@
 import torchvision
 
 sys.path.append("..")
-from profiler import Profiler
+from amanda_profiler import amandaProfiler
 
 def main():
 
@@ -16,7 +16,7 @@ def main():
 	metric = "KernelRoofline"
 	# Nvidia Geforce RTX 2080 Ti: 1350MHz, 13.45 Single-Precision TFlops, 616GB/s
 	supplyInfo = [1350, 13.45, 616]
-	profiler = Profiler(metric)
+	profiler = amandaProfiler(metric)
 	profiler.setConfigs(metric=metric, supplyInfo=supplyInfo)
 
 	with amanda.tool.apply(profiler.counter):

diff --git a/src/amanda/profiler/pytorch/test/opInfo_test.py b/src/amanda/profiler/pytorch/test/opInfo_test.py
@@ -5,7 +5,7 @@
 import torchvision
 
 sys.path.append("..")
-from profiler import Profiler
+from amanda_profiler import amandaProfiler
 
 def main():
 
@@ -15,7 +15,7 @@ def main():
 	x = torch.rand((32, 3, 227, 227)).to(device)
 
 	metric = "OpInfo"
-	profiler = Profiler(metric)
+	profiler = amandaProfiler(metric)
 	profiler.setConfigs(metric=metric, supplyInfo=[])
 
 	with amanda.tool.apply(profiler.counter):

diff --git a/src/amanda/profiler/pytorch/test/opRoofline_test.py b/src/amanda/profiler/pytorch/test/opRoofline_test.py
@@ -5,7 +5,7 @@
 import torchvision
 
 sys.path.append("..")
-from profiler import Profiler
+from amanda_profiler import amandaProfiler
 
 def main():
 
@@ -17,7 +17,7 @@ def main():
 	metric = "OpRoofline"
 	# Nvidia Geforce RTX 2080 Ti: 1350MHz, 13.45 Single-Precision TFlops, 616GB/s
 	supplyInfo = [1350, 13.45, 616]
-	profiler = Profiler(metric)
+	profiler = amandaProfiler(metric)
 	profiler.setConfigs(metric=metric, supplyInfo=supplyInfo)
 
 	with amanda.tool.apply(profiler.counter):

diff --git a/src/amanda/profiler/pytorch/torchMetrics.py b/src/amanda/profiler/pytorch/torchMetrics.py
@@ -41,7 +41,7 @@ def kernelInfo(opList, timeList, apiList, rtList, dataList):
 	res.to_csv("./Experiments/kernelInfo_result.csv", index=False, sep=',')
 
 
-# Now Information: OpIndex, OpName, TotalExecutionTime, MaxKernelIndex, MaxKernelName, MaxKernelExecutionTime, kernelNumTracer, kernelNumCounter, TotalDramRead, TotalDramWrite, TotalSFOp, TotalElapsedCycles
+# Now Information: OpIndex, OpName, TotalExecutionTime, MaxKernelIndex, MaxKernelName, MaxKernelExecutionTime, kernelNumTracer, kernelNumCounter, TotalElapsedCycles, TotalDramRead, TotalDramWrite, TotalSFOp
 def opInfo(opList, startTimeList, endTimeList, apiList, rtList, dataList):
 
 	# Information collected by tracer: OpIndex, OpName, TotalExecutionTime, MaxKernelIndex, MaxKernelName, MaxKernelExecutionTime, kernelNumTracer 

diff --git a/src/amanda/profiler/tensorflow/__init__.py b/src/amanda/profiler/tensorflow/__init__.py
@@ -1,3 +1,3 @@
 from amanda_tracer import *
 from amanda_counter import *
-from profiler import *
+from amanda_profiler import *
diff --git a/src/amanda/profiler/tensorflow/profiler.py → ...da/profiler/tensorflow/amanda_profiler.py b/src/amanda/profiler/tensorflow/profiler.py → ...da/profiler/tensorflow/amanda_profiler.py
@@ -6,8 +6,9 @@
 
 from utils import setConfigsMetric
 from tfMetrics import kernelInfo, opInfo
+from metrics import opInfoTracer, opInfoCounter, kernelInfoTracer, kernelInfoCounter
 
-class Profiler():
+class amandaProfiler():
 	def __init__(self, metric) -> None:
 		self.__metric = metric
 
@@ -45,7 +46,7 @@ def setConfigs(self, metric, supplyInfo, onlineOnly=False, offlineOnly=False):
 	def createSessionTracer(self):
 		gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.8)
 
-		config=tf.ConfigProto(gpu_options=gpu_options, intra_op_parallelism_threads=1, inter_op_parallelism_threads=1)
+		config=tf.ConfigProto(gpu_options=gpu_options, intra_op_parallelism_threads=0, inter_op_parallelism_threads=1)
 		session = tf.Session(config=config)
 		return session
 
@@ -57,25 +58,41 @@ def createSessionCounter(self):
 		return session
 
 	def showResults(self):
-		if self.__metric == "KernelInfo":
+		if self.__metric.find("KernelInfo") != -1:
+			self.tracer.activityFlushAll()
 			self.opListTracer = self.tracer.opList
 			self.opListCounter = self.counter.opList
 			self.startTimeList = self.tracer.getStartTimeLists()
 			self.traceDataRt = self.tracer.getTraceDataRt()
 			self.traceDataApi = self.tracer.getTraceDataApi()
 			self.countData = self.counter.getCountData()
-			kernelInfo(self.opListTracer, self.opListCounter, self.startTimeList, self.traceDataApi, self.traceDataRt, self.countData)
+			if self.__metric == "KernelInfo":
+				kernelInfo(self.opListTracer, self.opListCounter, self.startTimeList, self.traceDataApi, self.traceDataRt, self.countData)
+			elif self.__metric == "KernelInfoTracer":
+				kernelInfoTracer(self.opListTracer, self.startTimeList, self.traceDataApi, self.traceDataRt)
+			elif self.__metric == "KernelInfoCounter":
+				kernelInfoCounter(self.countData, flopCount=False)
+			else:
+				sys.exit("Profiler.Metric: " + self.__metric + " not supported")
 			return
 
-		if self.__metric == "OpInfo":
+		if self.__metric.find("OpInfo") != -1:
+			self.tracer.activityFlushAll()
 			self.opListTracer = self.tracer.opList
 			self.opListCounter = self.counter.opList
 			self.startTimeList = self.tracer.getStartTimeLists()
 			self.endTimeList = self.tracer.getEndTimeLists()
 			self.traceDataRt = self.tracer.getTraceDataRt()
 			self.traceDataApi = self.tracer.getTraceDataApi()
 			self.countData = self.counter.getCountData()
-			opInfo(self.opListTracer, self.opListCounter, self.startTimeList, self.endTimeList, self.traceDataApi, self.traceDataRt, self.countData)
+			if self.__metric == "OpInfo":
+				opInfo(self.opListTracer, self.opListCounter, self.startTimeList, self.endTimeList, self.traceDataApi, self.traceDataRt, self.countData)
+			elif self.__metric == "OpInfoTracer":
+				opInfoTracer(self.opListTracer, self.startTimeList, self.endTimeList, self.traceDataApi, self.traceDataRt)
+			elif self.__metric == "OpInfoCounter":
+				opInfoCounter(self.countData, flopCount=False)
+			else:
+				sys.exit("Profiler.Metric: " + self.__metric + " not supported")
 			return
 
 		sys.exit("Profiler.Metric: " + self.__metric + " not supported")

diff --git a/src/amanda/profiler/tensorflow/amanda_tracer.py b/src/amanda/profiler/tensorflow/amanda_tracer.py
@@ -38,7 +38,7 @@ def forward_instrumentation(self, context: amanda.OpContext):
 				if not tensor.dtype._is_ref_dtype
 			]
 
-		if len(op_outputs) != 0 and len(op_inputs) != 0:
+		if len(op_outputs) != 0 and len(op_inputs) != 0 and op.name.find("Initializer") == -1:
 		# if len(op_outputs) != 0 and len(op_inputs) != 0 and op.name.find("Relu") != -1:	
 			self.opList.append(op.name)
 			context.insert_before_op(
@@ -55,6 +55,7 @@ def forward_instrumentation(self, context: amanda.OpContext):
 
 	def init_trace(self, *inputs, op):
 		def extract_fn(*inputs):
+			self.activityFlushAll()
 			self.tracer.initTrace()
 			self.beforeCount += 1
 			return inputs
@@ -70,6 +71,7 @@ def extract_fn(*inputs):
 	def finish_trace(self, *outputs, op):	
 		def extract_fn(*outputs):
 			self.tracer.finishTrace()
+			self.tracer.activityFlushAll()
 			self.afterCount += 1
 			return outputs
 

diff --git a/src/amanda/profiler/tensorflow/test/kernelInfo_test.py b/src/amanda/profiler/tensorflow/test/kernelInfo_test.py
@@ -5,16 +5,16 @@
 from examples.common.tensorflow.model.resnet_50 import ResNet50
 
 sys.path.append("..")
-from profiler import Profiler
+from amanda_profiler import amandaProfiler
 
 def main():
 
 	model = ResNet50()
 	x = tf.random.uniform(shape = [1, 224, 224, 3])
 	y = model(x)
 
-	metric = "KernelInfo"
-	profiler = Profiler(metric)
+	metric = "KernelInfoTracer"
+	profiler = amandaProfiler(metric)
 	profiler.setConfigs(metric=metric, supplyInfo=[])
 
 	session_ = profiler.createSessionCounter()

diff --git a/src/amanda/profiler/tensorflow/test/opInfo_test.py b/src/amanda/profiler/tensorflow/test/opInfo_test.py
@@ -5,16 +5,16 @@
 from examples.common.tensorflow.model.resnet_50 import ResNet50
 
 sys.path.append("..")
-from profiler import Profiler
+from amanda_profiler import amandaProfiler
 
 def main():
 
 	model = ResNet50()
-	x = tf.random.uniform(shape = [1, 224, 224, 3])
+	x = tf.random.uniform(shape = [16, 224, 224, 3])
 	y = model(x)
 
 	metric = "OpInfo"
-	profiler = Profiler(metric)
+	profiler = amandaProfiler(metric)
 	profiler.setConfigs(metric=metric, supplyInfo=[])
 
 	session_ = profiler.createSessionCounter()

diff --git a/src/amanda/profiler/tracer/activity_trace_async.cpp b/src/amanda/profiler/tracer/activity_trace_async.cpp
@@ -792,16 +792,24 @@ void tracer::initTrace()
   // Some attributes require to be set before any CUDA context is created to be effective,
   // e.g. to be applied to all device buffer allocations (see documentation).
   CUPTI_CALL(cuptiActivityGetAttribute(CUPTI_ACTIVITY_ATTR_DEVICE_BUFFER_SIZE, &attrValueSize, &attrValue));
-  printf("%s = %llu\n", "CUPTI_ACTIVITY_ATTR_DEVICE_BUFFER_SIZE", (long long unsigned)attrValue);
+  // printf("%s = %llu\n", "CUPTI_ACTIVITY_ATTR_DEVICE_BUFFER_SIZE", (long long unsigned)attrValue);
+  attrValue *= 2;
+  CUPTI_CALL(cuptiActivitySetAttribute(CUPTI_ACTIVITY_ATTR_DEVICE_BUFFER_SIZE, &attrValueSize, &attrValue));
 
   CUPTI_CALL(cuptiActivityGetAttribute(CUPTI_ACTIVITY_ATTR_DEVICE_BUFFER_POOL_LIMIT, &attrValueSize, &attrValue));
-  printf("%s = %llu\n", "CUPTI_ACTIVITY_ATTR_DEVICE_BUFFER_POOL_LIMIT", (long long unsigned)attrValue);
+  // printf("%s = %llu\n", "CUPTI_ACTIVITY_ATTR_DEVICE_BUFFER_POOL_LIMIT", (long long unsigned)attrValue);
+  attrValue *= 2;
+  CUPTI_CALL(cuptiActivitySetAttribute(CUPTI_ACTIVITY_ATTR_DEVICE_BUFFER_POOL_LIMIT, &attrValueSize, &attrValue));
 
   CUPTI_CALL(cuptiActivityGetAttribute(CUPTI_ACTIVITY_ATTR_PROFILING_SEMAPHORE_POOL_SIZE, &attrValueSize, &attrValue));
-  printf("%s = %llu\n", "CUPTI_ACTIVITY_ATTR_DEVICE_BUFFER_POOL_SIZE", (long long unsigned)attrValue);
+  // printf("%s = %llu\n", "CUPTI_ACTIVITY_ATTR_DEVICE_BUFFER_POOL_SIZE", (long long unsigned)attrValue);
+  attrValue *= 2;
+  CUPTI_CALL(cuptiActivitySetAttribute(CUPTI_ACTIVITY_ATTR_DEVICE_BUFFER_POOL_LIMIT, &attrValueSize, &attrValue));
 
   CUPTI_CALL(cuptiActivityGetAttribute(CUPTI_ACTIVITY_ATTR_PROFILING_SEMAPHORE_POOL_LIMIT, &attrValueSize, &attrValue));
-  printf("%s = %llu\n", "CUPTI_ACTIVITY_ATTR_DEVICE_BUFFER_POOL_LIMIT", (long long unsigned)attrValue);
+  // printf("%s = %llu\n", "CUPTI_ACTIVITY_ATTR_DEVICE_BUFFER_POOL_LIMIT", (long long unsigned)attrValue);
+  attrValue *= 2;
+  CUPTI_CALL(cuptiActivitySetAttribute(CUPTI_ACTIVITY_ATTR_DEVICE_BUFFER_POOL_LIMIT, &attrValueSize, &attrValue));
 
   CUPTI_CALL(cuptiGetTimestamp(&startTimestamp));
   traceFile << "Update startTimeStamp: " << startTimestamp << " thread: " << pthread_self() << std::endl;