Merge pull request #95 from tingxingdong/develop

add batched gemm in google tester; add a ploting python script

clients/CMakeLists.txt

@@ -18,12 +18,12 @@ set( rocblas_clients_dependencies )
 
 # If the user does not explicitely specify BOOST_ROOT, build our
 # own on the fly
-#if( BUILD_CLIENTS_BENCHMARKS )
-#  if( NOT DEFINED BOOST_ROOT )
-#    include( external-boost )
-#    list( APPEND rocblas_clients_dependencies boost )
-#  endif( )
-#endif()
+if( BUILD_CLIENTS_BENCHMARKS )
+  if( NOT DEFINED BOOST_ROOT )
+    include( external-boost )
+    list( APPEND rocblas_clients_dependencies boost )
+  endif( )
+endif()
 
 if( BUILD_CLIENTS_TESTS )
   if( NOT DEFINED GTEST_ROOT )

clients/benchmarks/perf_script/README.txt

@@ -1,13 +1,16 @@
 
-Instruction to run the script to collect performance data
+Instruction to run the python script to collect performance data
 
-1) copy client.exe (Windows, client on Linux) to this folder 
+1) copy client.exe (Windows, client on Linux) to this folder
+   For example, "cp ../client ./" on linux
 
 2) "python measurePerformance.py --help" to see how to run the command line mode
     For example, "python measurePerformance.py -s 64-5760:64 -f gemm -r s --transa none --transb transpose > sgemm.txt" 
     will collect performance of sgemm NT of square matrices [64:5760] with step 64.
     All the output will be dumped into sgemm.txt
 
-3) Open the sgemm.txt and use "grep" to fetch the performance line. You can extract the Gflop/s number from the performance line
+3)  Plot the figure by refactoring with plotPerformance.py provided. 
+    "python plotPerformance.py" by default
+    will read step (2) output sgemm.txt and plot a performance figure in sgemm.pdf
 
 4) You can ingore the other automatic generated .txt files and folders during the run. 

clients/benchmarks/perf_script/plotPerformance.py

@@ -20,279 +20,33 @@
 import subprocess
 import itertools
 import os
-import matplotlib
+import matplotlib.pyplot as plt
 import pylab
 from matplotlib.backends.backend_pdf import PdfPages
-from blasPerformanceTesting import *
-
-def plotGraph(dataForAllPlots, title, plottype, plotkwargs, xaxislabel, yaxislabel):
-  """
-  display a pretty graph
-  """
-  colors = ['k','y','m','c','r','b','g']
-  #plottype = 'plot'
-  for thisPlot in dataForAllPlots:
-    getattr(pylab, plottype)(thisPlot.xdata, thisPlot.ydata,
-                             '{}.-'.format(colors.pop()), 
-                             label=thisPlot.label, **plotkwargs)
-  if len(dataForAllPlots) > 1:
-    pylab.legend(loc='best')
-
-  pylab.title(title)
-  pylab.xlabel(xaxislabel)
-  pylab.ylabel(yaxislabel)
-  pylab.grid(True)
-
-  if args.outputFilename == None:
-    # if no pdf output is requested, spit the graph to the screen . . .
-    pylab.show()
-  else:
-    pylab.savefig(args.outputFilename,dpi=(1024/8))
-    # . . . otherwise, gimme gimme pdf
-    #pdf = PdfPages(args.outputFilename)
-    #pdf.savefig()
-    #pdf.close()
-
-######## plotFromDataFile() Function to plot from data file begins ########
-def plotFromDataFile():
-  data = []
-  """
-  read in table(s) from file(s)
-  """
-  for thisFile in args.datafile:
-    if not os.path.isfile(thisFile):
-      print 'No file with the name \'{}\' exists. Please indicate another filename.'.format(thisFile)
-      quit()
-
-    results = open(thisFile, 'r')
-    results_contents = results.read()
-    results_contents = results_contents.rstrip().split('\n')
-
-    firstRow = results_contents.pop(0)
-    print firstRow
-    print blas_table_header()
-    print firstRow.rstrip()==blas_table_header()
-    if firstRow.rstrip() != blas_table_header():
-      print 'ERROR: input file \'{}\' does not match expected format.'.format(thisFile)
-      quit()
-
-    for row in results_contents:
-        row = row.split(',')
-        row = TableRow(BlasTestCombination(row[0],row[1],row[2],row[3],row[4],row[5],row[6],row[7],row[8],row[9],row[10],row[11],row[12],row[13],row[14], row[15], row[16], row[17][1:], row[17][0], row[18], row[19], row[20]), row[21])
-        data.append(BlasGraphPoint(row.parameters.sizem, row.parameters.sizen, row.parameters.sizek, row.parameters.lda, row.parameters.ldb, row.parameters.ldc, row.parameters.offa , row.parameters.offb , row.parameters.offc , row.parameters.device, row.parameters.order, row.parameters.transa, row.parameters.transb, row.parameters.precision + row.parameters.function, row.parameters.library, row.parameters.label, row.gflops))
-
-  """
-  data sanity check
-  """
-  # if multiple plotvalues have > 1 value among the data rows, the user must specify which to plot
-  multiplePlotValues = []
-  for option in plotvalues:
-    values = []
-    for point in data:
-      values.append(getattr(point, option)) 
-    multiplePlotValues.append(len(set(values)) > 1)
-  if multiplePlotValues.count(True) > 1 and args.plot == None:
-    print 'ERROR: more than one parameter of {} has multiple values. Please specify which parameter to plot with --plot'.format(plotvalues)
-    quit()
-
-  # if args.graphxaxis is not 'problemsize', the user should know that the results might be strange
-  #if args.graphxaxis != 'problemsize':
-  #  xaxisvalueSet = []
-  #  for option in xaxisvalues:
-  #    if option != 'problemsize':
-  #      values = []
-  #      for point in data:
-  #        values.append(getattr(point, option)) 
-  #      xaxisvalueSet.append(len(set(values)) > 1)
-  #  if xaxisvalueSet.count(True) > 1:
-  #    print 'WARNING: more than one parameter of {} is varied. unexpected results may occur. please double check your graphs for accuracy.'.format(xaxisvalues)
-
-  # multiple rows should not have the same input values
-  #pointInputs = []
-  #for point in data:
-  #  pointInputs.append(point.__str__().split(';')[0])
-  #if len(set(pointInputs)) != len(data):
-  #  print 'ERROR: imported table has duplicate rows with identical input parameters'
-  #  quit()
-
-  """
-  figure out if we have multiple plots on this graph (and what they should be)
-  """
-  if args.plot != None:
-    multiplePlots = args.plot
-  elif multiplePlotValues.count(True) == 1 and plotvalues[multiplePlotValues.index(True)] != 'sizek':
-    # we don't ever want to default to sizek, because it's probably going to vary for most plots
-    # we'll require the user to explicitly request multiple plots on sizek if necessary
-    multiplePlots = plotvalues[multiplePlotValues.index(True)]
-  else:
-    # default to device if none of the options to plot have multiple values
-    multiplePlots = 'device'
-
-  """
-  assemble data for the graphs
-  """
-  data.sort(key=lambda row: int(getattr(row, args.graphxaxis)))
-
-  # choose scale for x axis
-  if args.xaxisscale == None:
-    # user didn't specify. autodetect
-    if int(getattr(data[len(data)-1], args.graphxaxis)) > 2000: # big numbers on x-axis
-      args.xaxisscale = 'log2'
-    elif int(getattr(data[len(data)-1], args.graphxaxis)) > 10000: # bigger numbers on x-axis
-      args.xaxisscale = 'log10'
-    else: # small numbers on x-axis
-      args.xaxisscale = 'linear'
-
-  if args.xaxisscale == 'linear':
-    plotkwargs = {}
-    plottype = 'plot'
-  elif args.xaxisscale == 'log2':
-    plottype = 'semilogx'
-    plotkwargs = {'basex':2}
-  elif args.xaxisscale == 'log10':
-    plottype = 'semilogx'
-    plotkwargs = {'basex':10}
-  else:
-    print 'ERROR: invalid value for x-axis scale'
-    quit()
-
-  plots = set(getattr(row, multiplePlots) for row in data)
-
-  class DataForOnePlot:
-    def __init__(self, inlabel, inxdata, inydata):
-      self.label = inlabel
-      self.xdata = inxdata
-      self.ydata = inydata
-
-  dataForAllPlots = []
-  for plot in plots:
-    dataForThisPlot = itertools.ifilter( lambda x: getattr(x, multiplePlots) == plot, data)
-    dataForThisPlot = list(itertools.islice(dataForThisPlot, None))
-    #if args.graphxaxis == 'problemsize':
-    #  xdata = [int(row.x) * int(row.y) * int(row.z) * int(row.batchsize) for row in dataForThisPlot]
-    #else:
-    xdata = [getattr(row, args.graphxaxis) for row in dataForThisPlot]
-    ydata = [getattr(row, args.graphyaxis) for row in dataForThisPlot]
-    dataForAllPlots.append(DataForOnePlot(plot,xdata,ydata))
-
-  """
-  assemble labels for the graph or use the user-specified ones
-  """
-  if args.graphtitle:
-    # use the user selection
-    title = args.graphtitle
-  else:
-    # autogen a lovely title
-    title = 'Performance vs. ' + args.graphxaxis.capitalize()
-
-  if args.xaxislabel:
-    # use the user selection
-    xaxislabel = args.xaxislabel
-  else:
-    # autogen a lovely x-axis label
-    if args.graphxaxis == 'cachesize':
-      units = '(bytes)'
-    else:
-      units = '(datapoints)'
-
-    xaxislabel = args.graphxaxis + ' ' + units
-
-  if args.yaxislabel:
-    # use the user selection
-    yaxislabel = args.yaxislabel
-  else:
-    # autogen a lovely y-axis label
-    if args.graphyaxis == 'gflops':
-      units = 'GFLOPS'
-    yaxislabel = 'Performance (' + units + ')'
-
-  """
-  display a pretty graph
-  """
-  colors = ['k','y','m','c','r','b','g']
-
-  for thisPlot in dataForAllPlots:
-    getattr(pylab, plottype)(thisPlot.xdata, thisPlot.ydata, '{}.-'.format(colors.pop()), label=thisPlot.label, **plotkwargs)
-
-  if len(dataForAllPlots) > 1:
-    pylab.legend(loc='best')
-
-  pylab.title(title)
-  pylab.xlabel(xaxislabel)
-  pylab.ylabel(yaxislabel)
-  pylab.grid(True)
-
-  if args.outputFilename == None:
-    # if no pdf output is requested, spit the graph to the screen . . .
-    pylab.show()
-  else:
-    # . . . otherwise, gimme gimme pdf
-    #pdf = PdfPages(args.outputFilename)
-    #pdf.savefig()
-    #pdf.close()
-    pylab.savefig(args.outputFilename,dpi=(1024/8))
-######### plotFromDataFile() Function to plot from data file ends #########
-
-
-
-######## "main" program begins #####
-"""
-define and parse parameters
-"""
-xaxisvalues = ['sizem','sizen','sizek']
-yaxisvalues = ['gflops']
-plotvalues = ['lda','ldb','ldc','sizek','device','label','order','transa','transb','function','library']
-
-
-
-parser = argparse.ArgumentParser(description='Plot performance of the clblas\
-    library. clblas.plotPerformance.py reads in data tables from clblas.\
-    measurePerformance.py and plots their values')
-fileOrDb = parser.add_mutually_exclusive_group(required=True)
-fileOrDb.add_argument('-d', '--datafile',
-  dest='datafile', action='append', default=None, required=False,
-  help='indicate a file to use as input. must be in the format output by\
-  clblas.measurePerformance.py. may be used multiple times to indicate\
-  multiple input files. e.g., -d cypressOutput.txt -d caymanOutput.txt')
-parser.add_argument('-x', '--x_axis',
-  dest='graphxaxis', default=None, choices=xaxisvalues, required=True,
-  help='indicate which value will be represented on the x axis. problemsize\
-      is defined as x*y*z*batchsize')
-parser.add_argument('-y', '--y_axis',
-  dest='graphyaxis', default='gflops', choices=yaxisvalues,
-  help='indicate which value will be represented on the y axis')
-parser.add_argument('--plot',
-  dest='plot', default=None, choices=plotvalues,
-  help='indicate which of {} should be used to differentiate multiple plots.\
-      this will be chosen automatically if not specified'.format(plotvalues))
-parser.add_argument('--title',
-  dest='graphtitle', default=None,
-  help='the desired title for the graph generated by this execution. if\
-      GRAPHTITLE contains any spaces, it must be entered in \"double quotes\".\
-      if this option is not specified, the title will be autogenerated')
-parser.add_argument('--x_axis_label',
-  dest='xaxislabel', default=None,
-  help='the desired label for the graph\'s x-axis. if XAXISLABEL contains\
-      any spaces, it must be entered in \"double quotes\". if this option\
-      is not specified, the x-axis label will be autogenerated')
-parser.add_argument('--x_axis_scale',
-  dest='xaxisscale', default=None, choices=['linear','log2','log10'],
-  help='the desired scale for the graph\'s x-axis. if nothing is specified,\
-      it will be selected automatically')
-parser.add_argument('--y_axis_label',
-  dest='yaxislabel', default=None,
-  help='the desired label for the graph\'s y-axis. if YAXISLABEL contains any\
-      spaces, it must be entered in \"double quotes\". if this option is not\
-      specified, the y-axis label will be autogenerated')
-parser.add_argument('--outputfile',
-  dest='outputFilename', default=None,
-  help='name of the file to output graphs. Supported formats: emf, eps, pdf, png, ps, raw, rgba, svg, svgz.')
-
-args = parser.parse_args()
-
-if args.datafile != None:
-  plotFromDataFile()
-else:
-  print "Atleast specify if you want to use text files or database for plotting graphs. Use -h or --help option for more details"
-  quit()
 
+os.system( "grep NT sgemm.txt > sgemm_NT.csv" )
+input = open ('sgemm_NT.csv', 'r')
+x = []
+y = []
+shape = ''
+for line in input:
+    line = line.replace("(", ",")
+    line = line.replace(")", ",")
+    value = line.split(',')
+    x.append(value[1])
+    y.append(value[7])
+    shape = value[0]
+    #print value
+
+
+f = plt.figure()
+plt.rcParams.update({'font.size':20})
+plt.xlabel('M=N=K')
+plt.ylabel("Gflop/s")
+plt.title('rocBLAS SGEMM '  + shape)
+plt.yticks()
+plt.grid(True)
+plt.legend( loc = 2)
+plot1 = plt.plot(x, y)
+f.savefig("sgemm.pdf", bbox_inches='tight')
+input.close()

clients/gtest/CMakeLists.txt

@@ -84,6 +84,7 @@ find_package( HIP REQUIRED )
 if( BUILD_WITH_TENSILE )
 set(Tensile_TEST_SRC
     gemm_gtest.cpp
+    gemm_batched_gtest.cpp
     trsm_gtest.cpp
     )
 endif()

clients/gtest/gemm_batched_gtest.cpp

@@ -44,10 +44,10 @@ Representative sampling is sufficient, endless brute-force sampling is not neces
 const
 vector<vector<int>> matrix_size_range = {
                                         {-1, -1, -1, -1, 1, 1},
-                                        {10, 10, 20, 100, 10, 10},
-                                        {128, 128, 64, 128, 128, 128},
-                                        {600,500, 500, 500, 600, 500},
-                                        {1024, 1024, 1024, 1024, 1024, 1024}
+                                        {32, 32, 32, 100, 100, 100},
+                                        {64, 64, 64, 128, 128, 128},
+                                        {128, 128, 128, 128, 128, 128},
+                                        {500, 500, 500, 500, 600, 500},
                                        };
 
 
@@ -75,11 +75,11 @@ vector<vector<char>> transA_transB_range = {
 //number of gemms in batched gemm
 const
 vector<int> batch_count_range = {
-                                        {-1},
-                                        {0},
-                                        {1},
-                                        {10},
-                                        {100},
+                                        -1,
+                                        0,
+                                        1,
+                                        10,
+                                        100,
                                        };
 
 

clients/gtest/gemm_gtest.cpp

@@ -44,7 +44,7 @@ vector<vector<int>> matrix_size_range = {
 const
 vector<vector<int>> full_matrix_size_range = {
                                         {192, 192, 192, 192, 192, 192},
-                                        {640, 640, 640, 640, 640, 640},
+                                        {640, 640, 640, 960, 960, 960},
                                         {1000, 1000, 1000, 1000, 1000, 1000},
                                         {4011, 4011, 4011, 4011, 4011, 4011},
                                        };