dnn_analysis.py

#!/usr/bin/python2.7
import matplotlib
import matplotlib.pyplot as plt
import matplotlib.ticker as ticker
import matplotlib.patches as patches
from matplotlib.ticker import FuncFormatter
import numpy as np
import math

# This module is used for visualizing the result.


# This is a function to plot the utilization of buffer 
# and systolic array for a single DNN under one particular
# hardware configuration
def plot_util_dnn(res, suffix):
    # initialize the array from the result first.
    sa_util = []
    buf_util = []
    # comp_bound = {'x':[], 'y': []}
    # mem_bound = {'x':[], 'y': []}
    cnt = 1

    # the content in the result is listed as:
    for item in res:
        sa_util.append(item[2])
        buf_util.append(item[3])
        # if item[-1] == 'C':
        #     comp_bound['x'].append(cnt)
        #     comp_bound['y'].append(item[2]*item[3])
        # else:
        #     mem_bound['x'].append(cnt)
        #     mem_bound['y'].append(item[2]*item[3])
        # cnt += 1

    x_axis_ls = range(1, len(res)+1)
    plt.rc('font', size=10)
    ax1 = plt.figure(figsize=(8, 3)).add_subplot(111)
    ax1.set_ylabel('PE util', fontsize=14, fontweight='bold')
    plt.xticks(rotation=60)
    plt.setp(ax1.spines.values(), linewidth=2)

    ax2 = ax1.twinx()
    ax2.set_ylabel('Buf Util', fontsize=14, fontweight='bold')

    p1 = ax1.plot(x_axis_ls, sa_util, color='#71985E', linestyle=':', linewidth=2, \
                marker='d',markersize=8, markeredgewidth=1.5, markeredgecolor='k');
    
    p2 = ax2.plot(x_axis_ls, buf_util, color='#FFBF56', linestyle=':', linewidth=2, \
                    marker='o',markersize=8, markeredgewidth=1.5, markeredgecolor='k');

    # p3 = ax1.bar(comp_bound['x'], comp_bound['y'], 0.5, align='center',color='#8154D1',\
    #     edgecolor=['k']*len(x_axis_ls), linewidth=1.5, hatch="/");
    
    # p4 = ax2.bar(mem_bound['x'], mem_bound['y'], 0.5, align='center', \
    #     color='#5b87f2', edgecolor=['k']*len(x_axis_ls), linewidth=1.5, hatch="\\");
    # plt.subplots_adjust(left=0.1, bottom=0.25, right=0.9, top=0.9,
    #             wspace=0.2, hspace=0.2)
    plt.xticks(x_axis_ls, [ "Layer" + str(n) for n in x_axis_ls])
    ax1.set_ylim(0.0, 1.0)
    ax2.set_ylim(0.0, 1.0)
    ax1.tick_params(axis="y",direction="in")
    ax2.tick_params(axis="y",direction="in")
    ax1.tick_params(axis="x",direction="in")
    plt.grid(color='grey', which='major', axis='y', linestyle='--')
    ax1.yaxis.set_major_formatter(FuncFormatter(lambda y, _: '{:.0%}'.format(y))) 
    ax2.yaxis.set_major_formatter(FuncFormatter(lambda y, _: '{:.0%}'.format(y))) 
    plt.legend((p1[0], p2[0]), ('SA util', 'Buf util',),\
     bbox_to_anchor=(0., 1.01, 1., .101), loc=3,
           ncol=2, mode="expand", borderaxespad=0., frameon=False)
    ax1.set_axisbelow(True)
    
    plt.savefig(suffix+"_layer_util.pdf");
    print(sa_util)
    print(buf_util)


def profile_layer_cycle(res, suffix):
    # initialize the array from the result first.
    total_transfer = []
    total_cycle = []

    # the content in the result is listed as:
    for item in res:
        total_transfer.append(item[0])
        total_cycle.append(item[1])

    x_axis_ls = range(1, len(res)+1)
    plt.rc('font', size=10)
    ax1 = plt.figure(figsize=(8, 3)).add_subplot(111)
    ax1.set_ylabel('Norm. Energy', fontsize=14, fontweight='bold')
    plt.xticks(rotation=60)
    plt.setp(ax1.spines.values(), linewidth=2)

    ax2 = ax1.twinx()
    ax2.set_ylabel('Norm. Speedup', fontsize=14, fontweight='bold')

    p1 = ax1.plot(x_axis_ls,[float(i)/total_transfer[0] for i in total_transfer], color='#71985E', linestyle=':', linewidth=2, \
                marker='d',markersize=8, markeredgewidth=1.5, markeredgecolor='k');
    
    p2 = ax2.plot(x_axis_ls, [float(i)/total_cycle[0] for i in total_cycle], color='#FFBF56', linestyle=':', linewidth=2, \
                    marker='o',markersize=8, markeredgewidth=1.5, markeredgecolor='k');
    plt.subplots_adjust(left=0.1, bottom=0.25, right=0.9, top=0.9,
                wspace=0.2, hspace=0.2)
    # ax1.set_ylim(pow(10, 6), pow(10, 8))
    # ax2.set_ylim(pow(10, 6), pow(10, 8))
    ax1.set_yscale('log')
    ax2.set_yscale('log')
    plt.xticks(x_axis_ls, [ "Layer" + str(n) for n in x_axis_ls])
    ax1.tick_params(axis="y",direction="in")
    ax2.tick_params(axis="y",direction="in")
    ax1.tick_params(axis="x",direction="in")
    plt.grid(color='grey', which='major', axis='y', linestyle='--')
    plt.legend((p1[0], p2[0]), ('mem. traffic', 'cycle'), \
            bbox_to_anchor=(0., 1.01, 1., .101), loc=3,
            ncol=2, mode="expand", borderaxespad=0., frameon=False)
    ax1.set_axisbelow(True)
    
    plt.savefig(suffix+"_layer_cycle.pdf");
    print(total_transfer)
    print(total_cycle)



'''
The functions below are to profile the impacts of different bandwidth,
buffer size, and systolic array size on overall system.
'''
def plot_sa_size(res, low, high, step):

    plt.rc('font', size=10)
    ax1 = plt.figure(figsize=(16, 3)).add_subplot(111)
    ax1.set_ylabel('PE Util', fontsize=14, fontweight='bold')
    ax1.set_xlabel('Systolic Array Size', fontsize=14, fontweight='bold')
    plt.setp(ax1.spines.values(), linewidth=2)

    ax2 = ax1.twinx()
    ax2.set_ylabel('Buf Util', fontsize=14, fontweight='bold')
    p1 = ax1.bar([i - 0.3 for i in range(len(res['sa_avg']))], res['sa_avg'], \
        0.3, yerr=res['sa_std'], align='center',color='#8154D1',\
        edgecolor=['k']*len(res['sa_avg']), linewidth=1.5, hatch="/");
    
    p2 = ax2.bar(range(len(res['buf_avg'])), res['buf_avg'], 0.3, yerr=res['buf_std'], align='center', \
        color='#5b87f2', edgecolor=['k']*len(res['buf_avg']), linewidth=1.5, hatch="\\");
    plt.subplots_adjust(left=0.1, bottom=0.20, right=0.9, top=0.9,
                wspace=0.2, hspace=0.2)
    ax2.set_ylim(0.0, 1.0)
    ax1.set_ylim(0.0, 1.0)
    plt.xticks(range(len(res['sa_avg'])), range(low, high+step, step))
    plt.legend((p1[0], p2[0]), ('SA util', 'Buf util'), \
            bbox_to_anchor=(0., 1.01, 1., .101), loc=3, \
            ncol=2, mode="expand", borderaxespad=0., frameon=False)
    ax1.set_axisbelow(True)
    
    plt.savefig("profile_sa.pdf");

def plot_sa_cycle(res, low, high, step):

    plt.rc('font', size=10)
    ax1 = plt.figure(figsize=(6, 2.5)).add_subplot(111)
    ax1.set_ylabel('Norm. Speedup', fontsize=14, fontweight='bold')
    ax1.set_xlabel('Systolic Array Size', fontsize=14, fontweight='bold')
    plt.setp(ax1.spines.values(), linewidth=2)

    ax2 = ax1.twinx()
    ax1.tick_params(axis="both",direction="in",which="both")
    ax2.tick_params(axis="both",direction="in",which="both",right=False)
    ax2.set_ylabel('Norm. Energy', fontsize=14, fontweight='bold')
    p1 = ax1.bar([i - 0.3 for i in range(len(res['cycle_avg']))], \
        [res['cycle_avg'][0]/float(i) for i in res['cycle_avg']], \
        0.3, align='center',color='#71985E',\
        edgecolor=['k']*len(res['cycle_avg']), linewidth=1.5, hatch="/");
    p2 = ax2.bar(range(len(res['traffic_avg'])), \
        [float(i)/res['traffic_avg'][0] for i in res['traffic_avg']], 0.3, align='center', \
        color='#FFBF56', edgecolor=['k']*len(res['traffic_avg']), linewidth=1.5, hatch="\\");
    plt.subplots_adjust(left=0.1, bottom=0.20, right=0.9, top=0.9,
                wspace=0.2, hspace=0.2)
    ax1.set_yscale('log')
    # ax2.set_yscale('log')
    ax1.set_ylim(0.5,100)
    ax2.set_ylim(0, 1)
    # ax1.set_ylim(pow(10, math.floor(math.log10(min(res['cycle_avg'])))), \
    #                 pow(10, math.ceil(math.log10(max(res['cycle_avg'])))))
    # ax2.set_ylim(pow(10, math.floor(math.log10(min(res['traffic_avg'])))), \
    #                 pow(10, math.ceil(math.log10(max(res['traffic_avg'])))))
    
    # ax2.yaxis.set_major_formatter(FuncFormatter(lambda y, _: '{:.1f}'.format(y))) 
    # ax2.yaxis.set_minor_formatter(FuncFormatter(lambda y, _: '{:.1f}'.format(y))) 
    plt.xticks(range(len(res['sa_avg'])), range(low, high+step, step))
    
    ax2.set_yticks([0, 0.2, 0.4, 0.6, 0.8, 1.0])
    plt.legend((p1[0], p2[0]), ('speedup', 'Energy'), \
            bbox_to_anchor=(0., 1.01, 1., .101), loc=3, \
            ncol=2, mode="expand", borderaxespad=0., frameon=False)
    ax1.set_axisbelow(True)
    
    plt.savefig("profile_sa_cycle.pdf");


def plot_buf_size(res, low, high, step, base, scale):
    
    plt.rc('font', size=10)
    ax1 = plt.figure(figsize=(16, 3)).add_subplot(111)
    ax1.set_ylabel('PE Util', fontsize=14, fontweight='bold')
    ax1.set_xlabel('Buffer Size (MB)', fontsize=14, fontweight='bold')
    plt.setp(ax1.spines.values(), linewidth=2)

    ax2 = ax1.twinx()
    ax2.set_ylabel('Buf Util', fontsize=14, fontweight='bold')
    p1 = ax1.bar([i - 0.3 for i in range(len(res['sa_avg']))], res['sa_avg'],\
     0.3, yerr=res['sa_std'], align='center',color='#8154D1',\
        edgecolor=['k']*len(res['sa_avg']), linewidth=1.5, hatch="/");
    
    p2 = ax2.bar(range(len(res['buf_avg'])), res['buf_avg'], 0.3, yerr=res['buf_std'], align='center', \
        color='#5b87f2', edgecolor=['k']*len(res['buf_avg']), linewidth=1.5, hatch="\\");
    plt.subplots_adjust(left=0.1, bottom=0.20, right=0.9, top=0.9,
                wspace=0.2, hspace=0.2)
    ax2.set_ylim(0.0, 1.0)
    ax1.set_ylim(0.0, 1.0)
    # ax2.set_yscale('log')
    plt.xticks(range(len(res['sa_avg'])), [base*scale*i for i in range(low, high, step)])
    plt.legend((p1[0], p2[0]), ('SA util', 'Buf util'), \
            bbox_to_anchor=(0.1, 1.01, 0.8, .101), loc=3, \
            ncol=2, mode="expand", borderaxespad=0., frameon=False)
    ax1.set_axisbelow(True)
    
    plt.savefig("profile_buf.pdf");

def plot_buf_cycle(res, arr, scale):
    
    plt.rc('font', size=10)

    ax1 = plt.figure(figsize=(6, 2.5)).add_subplot(111)
    ax1.set_ylabel('Norm. Speedup', fontsize=14, fontweight='bold')
    ax1.set_xlabel('Buffer Size (MB)', fontsize=14, fontweight='bold')
    plt.setp(ax1.spines.values(), linewidth=2)

    ax2 = ax1.twinx()
    # plt.grid(color='grey', which='major', axis='y', linestyle='--')
    ax2.set_ylabel('Norm. Energy', fontsize=14, fontweight='bold')
    
    p1 = ax1.bar([i - 0.3 for i in range(len(res['cycle_avg']))],\
     [float(res['cycle_avg'][0])/float(i) for i in res['cycle_avg']],\
     0.3, align='center',color='#71985E', edgecolor=['k']*len(res['sa_avg']), \
     linewidth=1.5, hatch="/");
    
    p2 = ax2.bar(range(len(res['traffic_avg'])), \
        [float(i)/res['traffic_avg'][0] for i in res['traffic_avg']], 0.3, align='center', \
        color='#FFBF56', edgecolor=['k']*len(res['buf_avg']), linewidth=1.5, hatch="\\");
    
    plt.subplots_adjust(left=0.1, bottom=0.20, right=0.9, top=0.9,
                wspace=0.2, hspace=0.2)
    # ax1.set_yscale('log')
    # ax2.set_yscale('log')
    # ax1.set_ylim(pow(10, math.floor(math.log10(min(res['cycle_avg'])))), \
    #                 pow(10, math.ceil(math.log10(max(res['cycle_avg'])))))
    # ax2.set_ylim(pow(10, math.floor(math.log10(min(res['traffic_avg'])))), \
    #                 pow(10, math.ceil(math.log10(max(res['traffic_avg'])))))
    ax1.set_ylim(0.5, 1.5)
    ax2.set_ylim(0, 1)
    ax1.set_yticks([0.5, 0.7, 0.9, 1.1, 1.3, 1.5])
    ax2.set_yticks([0.0, 0.2, 0.4, 0.6, 0.8, 1.0])
    
    plt.xticks(range(len(res['cycle_avg'])), arr)
    plt.legend((p1[0], p2[0]), ('speedup', 'energy'), \
            bbox_to_anchor=(0.1, 1.01, 0.8, .101), loc=3, \
            ncol=2, mode="expand", borderaxespad=0., frameon=False)
    ax1.set_axisbelow(True)
    
    plt.savefig("profile_buf_cycle.pdf");



def plot_bw_size(res, low, high):
    
    plt.rc('font', size=10)
    ax1 = plt.figure(figsize=(16, 3)).add_subplot(111)
    ax1.set_ylabel('PE Util', fontsize=14, fontweight='bold')
    ax1.set_xlabel('Memory Bandwidth (B/cycle)', fontsize=14, fontweight='bold')
    plt.setp(ax1.spines.values(), linewidth=2)

    ax2 = ax1.twinx()
    ax2.set_ylabel('Buf Util', fontsize=14, fontweight='bold')
    p1 = ax1.bar([i - 0.3 for i in range(len(res['sa_avg']))], res['sa_avg'],\
     0.3, yerr=res['sa_std'], align='center',color='#8154D1',\
        edgecolor=['k']*len(res['sa_avg']), linewidth=1.5, hatch="/");
    
    p2 = ax2.bar(range(len(res['buf_avg'])), res['buf_avg'], 0.3, yerr=res['buf_std'], align='center', \
        color='#5b87f2', edgecolor=['k']*len(res['buf_avg']), linewidth=1.5, hatch="\\");
    plt.subplots_adjust(left=0.1, bottom=0.20, right=0.9, top=0.9,
                wspace=0.2, hspace=0.2)
    ax2.set_ylim(0.0, 1.0)
    ax1.set_ylim(0.0, 1.0)
    plt.xticks(range(len(res['sa_avg'])), [pow(2, i) for i in range(low, high)])
    plt.legend((p1[0], p2[0]), ('SA util', 'Buf util'), \
            bbox_to_anchor=(0., 1.01, 1., .101), loc=3, \
            ncol=4, mode="expand", borderaxespad=0., frameon=False)
    ax1.set_axisbelow(True)
    
    plt.savefig("profile_bw.pdf");

def plot_bw_cycle(res, low, high):
    
    plt.rc('font', size=10)
    ax1 = plt.figure(figsize=(16, 3)).add_subplot(111)
    ax1.set_ylabel('Total Cycle', fontsize=14, fontweight='bold')
    ax1.set_xlabel('Memory Bandwidth (B/cycle)', fontsize=14, fontweight='bold')
    plt.setp(ax1.spines.values(), linewidth=2)

    ax2 = ax1.twinx()
    ax2.set_ylabel('Total Traffic', fontsize=14, fontweight='bold')
    p1 = ax1.bar([i - 0.3 for i in range(len(res['cycle_avg']))], res['cycle_avg'],\
     0.3, align='center',color='#8154D1',\
        edgecolor=['k']*len(res['cycle_avg']), linewidth=1.5, hatch="/");
    
    p2 = ax2.bar(range(len(res['traffic_avg'])), res['traffic_avg'], 0.3, align='center', \
        color='#5b87f2', edgecolor=['k']*len(res['traffic_avg']), linewidth=1.5, hatch="\\");
    plt.subplots_adjust(left=0.1, bottom=0.20, right=0.9, top=0.9,
                wspace=0.2, hspace=0.2)
    ax1.set_yscale('log')
    ax2.set_yscale('log')
    ax1.set_ylim(pow(10, math.floor(math.log10(min(res['cycle_avg'])))), \
                    pow(10, math.ceil(math.log10(max(res['cycle_avg'])))))
    ax2.set_ylim(pow(10, math.floor(math.log10(min(res['traffic_avg'])))), \
                    pow(10, math.ceil(math.log10(max(res['traffic_avg'])))))
    plt.xticks(range(len(res['cycle_avg'])), [pow(2, i) for i in range(low, high)])
    plt.legend((p1[0], p2[0]), ('cycle', 'traffic'), \
            bbox_to_anchor=(0.1, 1.01, 0.9, .101), loc=3, \
            ncol=4, mode="expand", borderaxespad=0., frameon=False)
    ax1.set_axisbelow(True)
    
    plt.savefig("profile_bw_cycle.pdf");