2023¶

可做的内容(NPU支持)
- Pytorch官方库：torchao、torchtune、torchtitan
- Pytorch官方特性：原生支持并行能力，Dtensor
- Pytorch三方库
困境
- 新特性/库的代码不稳定，现在支持之后变动可能会很大
- 由于人手和机器不足，余下的三方库很多，但是支持的必要性有待验证。客户的呼声不明，没有特别急用的库(之前适配过了)。
- 提前跟进库的适配，如果跟进过多，会导致后续持续看护库的人员压力会很大。

2023年10月16日
分类于 toLearn
需要 1 分钟阅读时间

Mkdocs

导言

mkdocs在今年支持了blog的基本功能，而且已经有探路者实践过了¹。也是时候升级博客生成器了。

2023年10月16日
分类于 Values
需要 1 分钟阅读时间

Research outline

读论文

读论文，要多篇，提炼overview 抓住立足点，创新点和展望

2023年10月16日
分类于 Algorithms
需要 1 分钟阅读时间

Concurrent-AMAT, C-AMAT

并发式平均存储访问时间模型

APC

侧重于测量方法的研究，给出了并发存储的测量方法和尺度。

在 APC 中，周期是存储活动周期 (memory active cycle)，不是通用的 CPU 周期，所以 APC 也叫 APMAC（存储活动周期平均访问数， access per memory active cycle）。

同时 APC 采用重叠 (overlapping) 的访存时间统计方法：在有两个或多个存储访问同时进行时，周期只增加一次。

2023年10月16日
分类于 Algorithms
需要 2 分钟阅读时间

HPL.dat file detailed explanation

HPL.dat

HPLinpack benchmark input file
Innovative Computing Laboratory, University of Tennessee
HPL.out      output file name (if any) 输出文件名
6            device out (6=stdout,7=stderr,file)

1            # of problems sizes (N) = sqrt((Memory Size in Gbytes * 1024 * 1024 * 1024) /8) * ratio
11136         Ns 矩阵规模

1            # of NBs block sizes (64~512)In the [96,104,112,120,128, …, 256] range; the multiple of 64
96           NBs 矩阵分块方法

0            PMAP process mapping (0=Row-,1=Column-major) 选择处理器阵列是按列的排列方式还是按行的排列方式。

1            # of process grids (P x Q)
2            Ps # two-dimensional block-cyclic data distribution = amount of processes; P≤Q
2            Qs 二维处理器网格（P×Q）

16.0         threshold 阈值

1            # of panel fact # 后面是L分解的方式
2            PFACTs (0=left, 1=Crout, 2=Right)
1            # of recursive stopping criterium
4            NBMINs (>= 1)
1            # of panels in recursion
2            NDIVs
1            # of recursive panel fact.
1            RFACTs (0=left, 1=Crout, 2=Right)
1            # of broadcast
1            BCASTs (0=1rg,1=1rM,2=2rg,3=2rM,4=Lng,5=LnM)
1            # of lookahead depth
1            DEPTHs (>=0)
2            SWAP (0=bin-exch,1=long,2=mix)
64           swapping threshold
0            L1 in (0=transposed,1=no-transposed) form
0            U  in (0=transposed,1=no-transposed) form
1            Equilibration (0=no,1=yes)
8            memory alignment in double (> 0)

合适的参数参考

2023年10月16日
分类于 Tutorials
需要 8 分钟阅读时间

Python Graph & Visualization

matplotlib VS plotly

matplotlib can use custom color and line style
plotly is more easily to quickly built up.

基础语法与概念

线性颜色柱的选择

https://plotly.com/python/builtin-colorscales/

same in matplotlib

plt.show 转发图像到本地

使用Dash: A web application framework for your data., 默认部署在localhost:8050端口

本地机器打通ssh隧道

ssh -L 8050:127.0.0.1:8050 -vN -f -l shaojiemike 202.38.72.23

科研画图

In scientific research plotting, it's always bar charts instead of line charts.

Possible reasons:

The visual footprint of point lines is relatively small compared to bar
Line charts are aggregated together when there is a meaningful comparison of data along the y-axis.

global setting

font things, Attention, global settings have higher priority to get work

matplotlib.rcParams.update({
    "pgf.texsystem": "pdflatex",
    'font.family': 'serif',
    # 'text.usetex': True, # comment to support bold font in legend, and font will be bolder
    # 'pgf.rcfonts': False,
})

layout

picture size

figure size (adjust x,y tick distance)

# mpl
fig.set_size_inches(w= 0.5 * x_count * (group_count+1.5), h=5.75 * 0.8) #(8, 6.5)

# plotly
fig.update_layout(height=350, width=50 * graph_data.size(),
                  margin=dict(b=10, t=10, l=20, r=5),
                  bargap=0.2 # x tick distance, 1 is the normalize-distance of adjacent bars
                  )

relative position

# mpl: Adjust the left margin to make room for the legend, left & right chart vertical line position from [0,1]
plt.subplots_adjust(left=0.1, right=0.8)

set x axis

# mpl:
x = np.arange(len(graph_data.x))  # the label locations, [0, 1, 2, 3, 4, 5, 6, 7]
# set the bar move to arg1 with name arg2
ax.set_xticks(x + (group_count-1)*0.5*width, graph_data.x)
plt.xticks(fontsize=graph_data.fontsize+4)
# Adjust x-axis limits to narrow the gap
plt.xlim(-(0.5+gap_count)*width, 
        x_count - 1 + (group_count-1)*width + (0.5+gap_count)*width)


# plotly

set y axis

vertical grid line
dick size
and range size

# mpl
ax.set_ylabel(yaxis_title, 
              fontsize=graph_data.fontsize,
              fontweight='bold')
plt.grid(True, which='major',axis='y', zorder=-1.0) # line and bar seems need to set zorder=10 to cover it
plt.yticks(np.arange(0, max_y ,10), fontsize=graph_data.fontsize) # step 10
plt.yscale('log',base=10) # or plt.yscale('linear')
ax.set_ylim(0.1, max_y)
## highlight selected y-label: https://stackoverflow.com/questions/73597796/make-one-y-axis-label-bold-in-matplotlib

# plotly
fig.update_layout(
      yaxis_range=[0,maxY],                   
      yaxis=dict(
          rangemode='tozero',  # Set the y-axis rangemode to 'tozero'
          dtick=maxY/10,
          gridcolor='rgb(196, 196, 196)', # grey
          gridwidth=1,
      ),
)

Bolden Contour Lines

Entire Figure

# mpl ?

# plotly: Add a rectangle shape to cover the entire subplot area
fig.add_shape(type="rect",
            xref="paper", yref="paper",
            x0=0, y0=0,
            x1=1, y1=1,
            line=dict(color="black", width=0.7))

and Each Bar Within

# mpl: Create a bar chart with bold outlines
plt.bar(categories, values, edgecolor='black', linewidth=2)

# plotly: ?

legend box

# mpl: 
ax.legend(loc='upper left', ncols=3, fontsize=graph_data.fontsize)
# legend out-of-figure, (1.02, 1) means anchor is upper right corner
plt.legend(loc='upper left', bbox_to_anchor=(1.02, 1), borderaxespad=0)

# Calculate the legend width based on the figure width
fig_size = plt.gcf().get_size_inches()
fig_width = fig_size[0]
# Move the legend to the center above the ceiling line
plt.legend(loc='upper center', bbox_to_anchor=(0.5, 1.1), 
  ncol=2,  # ncol=2 to have labels in one line
  frameon=False,  # frameon=False removes the legend box outline
  columnspacing=fig_width, # distance between each label
  handlelength=1.0, # label-box width (unit is text fontsize)
  handleheight=1.0, # label-box heigh (unit is text fontsize)
  prop={'size': 20, 'weight': 'bold'} # text fontsize          
)

bar

# mpl: white hugo hatch with black bar edge.
# Problem: because the bug of mpl. hatch color follow the edge color
# Solved: draw bar twice, first the white hugo hatch with white bar edge. Second empty figure with black bar edge.
# white doubel ref: https://stackoverflow.com/questions/38168948/how-to-decouple-hatch-and-edge-color-in-matplotlib
# ref2: https://stackoverflow.com/questions/71424924/how-to-change-the-edge-color-of-markers-patches-in-matplotlib-legend
color_palette = ['black', (193/255, 1/255, 1/255), 'w', (127/255, 126/255, 127/255), 'blue']
pattern_list = ["", "/", "+", "\\", "x"]
edgecolor_list = ['w', 'w', (0/255, 176/255, 80/255), 'w', 'w']

ax.bar(x + offset, measurement, width, label=species_name,
                       color=color_palette[idx],
                        hatch = pattern_list[idx],
                        edgecolor=edgecolor_list[idx],
                        linewidth=1,
                     )
ax.bar(x + offset, measurement, width, label=species_name,
            color = "none",
            edgecolor='black', linewidth=1,
              )
# related legend: https://stackoverflow.com/questions/71424924/how-to-change-the-edge-color-of-markers-patches-in-matplotlib-legend
handles1, labels1 = ax.get_legend_handles_labels()
plt.legend([handles1[2*idx]+handles1[2*idx+1] for idx in range(group_count)], 
            [labels1[2*idx] for idx in range(group_count)],  
            loc='upper center', bbox_to_anchor=(0.5, 1.12), 
            ncol=group_count,  # have labels in one line
            frameon=False,
            # bbox_transform=plt.gcf().transFigure,
            columnspacing=legend_width, 
            # handlelength=1.0, 
            handleheight=1.2, 
            prop={'size': graph_data.fontsize,
                'weight': 'heavy'}
) 


# plotly: find the overflow
overflow_pattern = ["/" if y > maxY  else "" for y in entry[1]]
fig.add_bar(x=x,y=yList, 
    name=barName, 
    marker=dict(
        color=color_list[i],
        pattern_shape = overflow_pattern,
        line=dict(color='black', width=2)
        ),
    textfont=dict(size=graph_data.fontsize),
)
# legend 
legend_num = len( barDict.items())
fig.update_layout(barmode="relative", 

        # legend_title="Legend Title",
        legend = dict(
          entrywidthmode='fraction', # https://plotly.com/python/legend/
          entrywidth= 0.2,
          x=0.5 - 0.5 * legend_num * 0.2,        # Set x to 0.5 for the center
          y=1.2,       # Set y to a value greater than 1 to move it above the plot
          orientation="h",  # Display legend items in a single line
        ),
)

Out-box text

To draw symmetry chart, we need to special highlight the overflow bar number.

If the ancher point locate in the plot box, it's easy to show text above the ceil line using textposition="bottom" like option. In the opposite scenario, plotly and mathplotlib all will hide the out-box text.

# plotly
fig.add_annotation(
            x=[x[0][i],x[1][i]],  # 注释的 x 坐标为 "bc"
            y=min(maxY,entry),  # 注释的 y 坐标为该列的最大值
            text=f"{entry:.2f}",  # 注释的文本内容
            # valign = "bottom", # text position in text box(default invisible)
            yanchor = "bottom", # text box position relative to anchor
            showarrow=False,  # 显示箭头
            # bgcolor="rgba(255, 255, 255, 0.8)",  # 注释框背景颜色
            font=dict(size=graph_data.fontsize+2)  # 注释文本字体大小
        )

# mathplotlib
# Create labels for overflowed values
for i, value in enumerate(values):
    if value > maxY:
        ax.annotate(f'Overflow: {value:.2f}', (i, maxY), ha='center', va='bottom', fontsize=12)

But mlb can write text out box.

ax.text(1, -1.6, 'Increasing', ha="center")

# first parameter is text, xy is the anchor point, xytext is the text, 
# xytext 2 xy is a relative distance
ax.annotate('yahaha', xy=(0, -0.1), xycoords='axes fraction', xytext=(1, -0.1))

out-box line

mathplotlib(mpl) can achieve this using ref, but there are few blogs about plotly.

# mpl: from 1*1 size full-graph (0.5,0.2) to point (1,0.8)
# transform=gcf().transFigure : gcf() stands for "get current figure," and .transFigure indicates that the coordinates provided in [0.5, 0.5], [0, 1] are in figure-relative coordinates. This means that the line's position is defined relative to the entire figure, not just the axes
# clip_on=False : This setting means that the line is not clipped at the edges of the axes. It allows the line to extend beyond the axes' boundaries.
from pylab import *  
plot([0.5, 1], [0.2, 0.8], color='lightgreen', linestyle='--', lw=1 ,transform=gcf().transFigure, clip_on=False)

# mpl: arrow from xy 2 xytext
# xycoords='figure fraction' to Add annotation to the figure (full graph)
ax.annotate('', xy=(0, -0.1), xycoords='axes fraction', xytext=(1, -0.1),\
arrowprops=dict(arrowstyle="->", color='violet'))

in-box line

# mpl:
ax.axhline(y=12, color='red', linestyle='--', label='Horizontal Line at y=12')
ax.axvline(x=3, color='green', linestyle='-.', label='Vertical Line at x=3')

# plotly ref: https://plotly.com/python/horizontal-vertical-shapes/
fig.add_vline(x=2.5, line_width=3, line_dash="dash", line_color="green") # dot
fig.add_hline(y=0.9)

实践

气泡图

二维无向图

教程

3D图

dash + plotly

如果防火墙是关闭的，你可以直接部署在external address上。使用docker也是可行的办法

app.run_server(debug=True, host='202.38.72.23')

3D 散点图，拟合曲面与网格图

实际案例

折线图

import matplotlib.pyplot as plt

X = [str(i) for i in metricValue]
Y = accuracyResult

# 设置图片大小
fig, ax = plt.subplots(figsize=(10, 6))  # 指定宽度为10英寸，高度为6英寸

plt.plot(X, Y, marker='o')
plt.xlabel('Threshold Percentage(%)', fontsize=12)  # 设置x轴标签字体大小为12
plt.ylabel('Average Execution Time of Static Method', fontsize=12)  # 设置y轴标签字体大小为12
plt.title('Tuning load store pressure', fontsize=14)  # 设置标题字体大小为14

for i in range(len(X)):
    plt.text(X[i], Y[i], Y[i], 
        fontsize=10, # 设置文本字体大小为10
        ha='center', # 设置水平对齐方式
        va='bottom')  # 设置垂直对齐方式

# 保存图片
plt.savefig(glv._get("resultPath") + f"tuning/{tuningLabel}/loadStorePressure.png", dpi=300)  # 设置dpi为300，可调整保存图片的分辨率

plt.show()  # 显示图片
plt.close()

fig = go.Figure()

 # color from https://stackoverflow.com/questions/68596628/change-colors-in-100-stacked-barchart-plotly-python
 color_list = ['rgb(29, 105, 150)', \
             'rgb(56, 166, 165)', \
    'rgb(15, 133, 84)',\
          'rgb(95, 70, 144)']
 sumList = [0 for i in range(len(x[0]))]
 for i, entry in enumerate( barDict.items()):
  barName=entry[0]
  yList = entry[1]
  ic(sumList,yList)
  sumList = [x + y for x, y in zip(yList, sumList)]
  fig.add_bar(x=x,y=yList, 
              name=barName, 
              text =[f'{val:.2f}' for val in yList], 
              textposition='inside',
              marker=dict(color=color_list[i]),
              error_y=dict(
    type='data',
    symmetric=False,
    color='purple',
    array=[i-j for i,j in zip(up_error,sumList)],
    arrayminus=[i-j for i,j in zip(sumList,down_error)],
       thickness=2, width=10),
              textfont=dict(size=8)
  )

en	cn
Knowledgeable	知己知彼 Be Skeptical
Independent	自己拥有研究
Smart	透过观象看本质
Soft	三人行必有我师