In [11]:
%%R -w 177 -h 89 -u mm -r 100
if(PDF_OUTPUT_DIR != "") {
outputpdf=paste(PDF_OUTPUT_DIR, "nr-table.pdf", sep="/")
pdf(outputpdf, width=7, height=3.5)
}
# Create a new session with the adequate data
S = new_session(bench, architecture)
# Clean the feature list
features = S$features[feature_set]
features = scale_features(features)
# Compute the distance in the feature space between every pair of codelets
distances = distancematrix(features)
# Compute a hierarchical clustering dendrogram
fit <- hclust(distances, method="ward")
fit = as.dendrogram(fit)
# Run the clustering and prediction method
clusters = cluster_codelets(S, feature_set, number_of_clusters)
prediction = cycle_prediction(S, feature_set, clusters)
prediction$Vec.Ratio.= round(S$features$Vec..ratio......all..FP)
# Read high-level classification data
hlclass = read.csv(paste(DATA_PATH,"hlclassification.csv",sep="/"), header=T, sep=",")
prediction = merge(prediction, hlclass, by="CodeletName")
# Reorder table prediction using the dendrogram codelet order
newprediction = data.frame()
labels <- prediction$CodeletName
for (i in labels[order.dendrogram(fit)]) {
newprediction = rbind(prediction[prediction$Codelet == i,], newprediction)
}
# Reorder cluster numbers by dendrogram
newprediction$Cluster = as.integer(factor(as.character(newprediction$Cluster),
levels=unique(as.character(newprediction$Cluster))))
prediction = newprediction
# Internally times (and therefore accelerations) are relative to machine cycles
# convert to absolute time using the architecture clocks frequencies
freq_ratio = S$ref_clk / S$tar_clk
prediction[,"realSpeedup"] <- 1/(prediction[,"realSpeedup"] * (freq_ratio))
prediction[,"Speedup"] = 1/(prediction[,"Speedup"] * (freq_ratio))
# Put acceleration of representatives between < >
isrepr = prediction[,"is.representative"] == "TRUE"
prediction[!isrepr,"realSpeedup"] = sprintf("%.4s",prediction[!isrepr,"realSpeedup"])
prediction[isrepr,"realSpeedup"] = sprintf("<%.4s>",prediction[isrepr,"realSpeedup"])
# Clean the codelet names suffixes
cleanCodeletNames <- function(D) {
for (suffix in c("_dp_sse", "_bet1_dt0_sse_initbet1", "_sp_sse", "_square_sp_sse", "_mp_sse"))
{
D$CodeletName = factor(sub(suffix, "", D$CodeletName))
}
D$CodeletName = factor(sub("hqr_12_square", "hqr_12_sq", D$CodeletName))
return(D)
}
prediction = cleanCodeletNames(prediction)
# Summarize codelet data by cluster number
D = ddply(prediction, .(Cluster), summarize,
Codelet=CodeletName,
"Computation Pattern"=Algorithm,
Stride=Array.Access..Stride,
Vec.=Vectorization,
"Vec. %" = Vec.Ratio.,
s=realSpeedup)
# Rename column Cluster to C
colnames(D)[1] = "C"
# Clean the D table so identical clusters in successive lines
# are not repeated
cleanf <- function(x){
oldx <- c(FALSE, x[-1]==x[-length(x)])
res <- x
res[oldx] <- ""
res
}
D[c("C")] <- lapply(D[c("C")], cleanf)
# Compute cluster separations row numbers
seps = as.list(which(D$C != "")-1)
# Plot the dendrogram and the table
par(mar=c(0,0,0,0), no.readonly=TRUE)
plot.new()
gl <- grid.layout(nrow=1, ncol=2, widths=unit(c(1,5), 'null'), heights=unit(c(1), 'null'))
vp.1 <- viewport(layout.pos.col=1, layout.pos.row=1)
vp.2 <- viewport(layout.pos.col=2, layout.pos.row=1)
pushViewport(viewport(layout=gl))
pushViewport(vp.1)
# Plot Dendrogram
p = ggdendrogram(fit, rotate=TRUE, labels=F)
p = p + theme(plot.margin = unit(c(0.2,-1.38,-0.94,0), "cm"))
p = p + geom_hline(yintercept=4.5, linetype="dashed", color="darkred")
p = p + annotate("text", size=3, y=14, x=1, color="darkred", label="cut for K = 14")
print(p, newpage=F)
popViewport()
pushViewport(vp.2)
# Plot Table
grid.table(D, gp =gpar(fontsize=6.5, lwd=1), show.rownames=F, padding.v = unit(1.2, "mm"),
padding.h = unit(3.4, "mm"))
# Plot Separators
for (s in seps[-1]) {
ss = nrow(D)-s
y = unit(1,"mm")+unit(ss*2.985,"mm")
grid.segments(0.058, y, 1-0.058, y)
}
popViewport()
if(PDF_OUTPUT_DIR != "") { dev.off(); print(paste("Saved to", outputpdf)) }
