I'm trying to speed up the solution time for a dynamic programming problem in Julia (v. 0.5.0), via parallel processing. The problem involves choosing the optimal values for every element of a 1073 x 19 matrix at every iteration, until successive matrix differences fall within a tolerance. I thought that, within each iteration, filling in the values for each element of the matrix could be parallelized. However, I'm seeing a huge performance degradation using SharedArray, and I'm wondering if there's a better way to approach parallel processing for this problem.
I construct the arguments for the function below:
est_params = [.788,.288,.0034,.1519,.1615,.0041,.0077,.2,0.005,.7196]
r = 0.015
tau = 0.35
rho =est_params[1]
sigma =est_params[2]
delta = 0.15
gamma =est_params[3]
a_capital =est_params[4]
lambda1 =est_params[5]
lambda2 =est_params[6]
s =est_params[7]
theta =est_params[8]
mu =est_params[9]
p_bar_k_ss =est_params[10]
beta = (1+r)^(-1)
sigma_range = 4
gz = 19
gp = 29
gk = 37
lnz=collect(linspace(-sigma_range*sigma,sigma_range*sigma,gz))
z=exp(lnz)
gk_m = fld(gk,2)
# Need to add mu somewhere to k_ss
k_ss = (theta*(1-tau)/(r+delta))^(1/(1-theta))
k=cat(1,map(i->k_ss*((1-delta)^i),collect(1:gk_m)),map(i->k_ss/((1-delta)^i),collect(1:gk_m)))
insert!(k,gk_m+1,k_ss)
sort!(k)
p_bar=p_bar_k_ss*k_ss
p = collect(linspace(-p_bar/2,p_bar,gp))
#Tauchen
N = length(z)
Z = zeros(N,1)
Zprob = zeros(Float32,N,N)
Z[N] = lnz[length(z)]
Z[1] = lnz[1]
zstep = (Z[N] - Z[1]) / (N - 1)
for i=2:(N-1)
Z[i] = Z[1] + zstep * (i - 1)
end
for a = 1 : N
for b = 1 : N
if b == 1
Zprob[a,b] = 0.5*erfc(-((Z[1] - mu - rho * Z[a] + zstep / 2) / sigma)/sqrt(2))
elseif b == N
Zprob[a,b] = 1 - 0.5*erfc(-((Z[N] - mu - rho * Z[a] - zstep / 2) / sigma)/sqrt(2))
else
Zprob[a,b] = 0.5*erfc(-((Z[b] - mu - rho * Z[a] + zstep / 2) / sigma)/sqrt(2)) -
0.5*erfc(-((Z[b] - mu - rho * Z[a] - zstep / 2) / sigma)/sqrt(2))
end
end
end
# Collecting tauchen results in a 2 element array of linspace and array; [2] gets array
# Zprob=collect(tauchen(gz, rho, sigma, mu, sigma_range))[2]
Zcumprob=zeros(Float32,gz,gz)
# 2 in cumsum! denotes the 2nd dimension, i.e. columns
cumsum!(Zcumprob, Zprob,2)
gm = gk * gp
control=zeros(gm,2)
for i=1:gk
control[(1+gp*(i-1)):(gp*i),1]=fill(k[i],(gp,1))
control[(1+gp*(i-1)):(gp*i),2]=p
end
endog=copy(control)
E=Array(Float32,gm,gm,gz)
for h=1:gm
for m=1:gm
for j=1:gz
# set the nonzero net debt indicator
if endog[h,2]<0
p_ind=1
else
p_ind=0
end
# set the investment indicator
if (control[m,1]-(1-delta)*endog[h,1])!=0
i_ind=1
else
i_ind=0
end
E[m,h,j] = (1-tau)*z[j]*(endog[h,1]^theta) + control[m,2]-endog[h,2]*(1+r*(1-tau)) +
delta*endog[h,1]*tau-(control[m,1]-(1-delta)*endog[h,1]) -
(i_ind*gamma*endog[h,1]+endog[h,1]*(a_capital/2)*(((control[m,1]-(1-delta)*endog[h,1])/endog[h,1])^2)) +
s*endog[h,2]*p_ind
elem = E[m,h,j]
if E[m,h,j]<0
E[m,h,j]=elem+lambda1*elem-.5*lambda2*elem^2
else
E[m,h,j]=elem
end
end
end
end
I then constructed the function with serial processing. The two for loops iterate through each element to find the largest value in a 1072-sized (=the gm scalar argument in the function) array:
function dynam_serial(E,gm,gz,beta,Zprob)
v = Array(Float32,gm,gz )
fill!(v,E[cld(gm,2),cld(gm,2),cld(gz,2)])
Tv = Array(Float32,gm,gz)
# Set parameters for the loop
convcrit = 0.0001 # chosen convergence criterion
diff = 1 # arbitrary initial value greater than convcrit
while diff>convcrit
exp_v=v*Zprob'
for h=1:gm
for j=1:gz
Tv[h,j]=findmax(E[:,h,j] + beta*exp_v[:,j])[1]
end
end
diff = maxabs(Tv - v)
v=copy(Tv)
end
end
Timing this, I get:
#time dynam_serial(E,gm,gz,beta,Zprob)
> 106.880008 seconds (91.70 M allocations: 203.233 GB, 15.22% gc time)
Now, I try using Shared Arrays to benefit from parallel processing. Note that I reconfigured the iteration so that I only have one for loop, rather than two. I also use v=deepcopy(Tv); otherwise, v is copied as an Array object, rather than a SharedArray:
function dynam_parallel(E,gm,gz,beta,Zprob)
v = SharedArray(Float32,(gm,gz),init = S -> S[Base.localindexes(S)] = myid() )
fill!(v,E[cld(gm,2),cld(gm,2),cld(gz,2)])
# Set parameters for the loop
convcrit = 0.0001 # chosen convergence criterion
diff = 1 # arbitrary initial value greater than convcrit
while diff>convcrit
exp_v=v*Zprob'
Tv = SharedArray(Float32,gm,gz,init = S -> S[Base.localindexes(S)] = myid() )
#sync #parallel for hj=1:(gm*gz)
j=cld(hj,gm)
h=mod(hj,gm)
if h==0;h=gm;end;
#async Tv[h,j]=findmax(E[:,h,j] + beta*exp_v[:,j])[1]
end
diff = maxabs(Tv - v)
v=deepcopy(Tv)
end
end
Timing the parallel version; and using a 4-core 2.5 GHz I7 processor with 16GB of memory, I get:
addprocs(3)
#time dynam_parallel(E,gm,gz,beta,Zprob)
> 164.237208 seconds (2.64 M allocations: 201.812 MB, 0.04% gc time)
Am I doing something incorrect here? Or is there a better way to approach parallel processing in Julia for this particular problem? I've considered using Distributed Arrays, but it's difficult for me to see how to apply them to the present problem.
UPDATE:
Per #DanGetz and his helpful comments, I turned instead to trying to speed up the serial processing version. I was able to get performance down to 53.469780 seconds (67.36 M allocations: 103.419 GiB, 19.12% gc time) through:
1) Upgrading to 0.6.0 (saved about 25 seconds), which includes the helpful #views macro.
2) Preallocating the main array I'm trying to fill in (Tv), per the section on Preallocating Outputs in the Julia Performance Tips: https://docs.julialang.org/en/latest/manual/performance-tips/. (saved another 25 or so seconds)
The biggest remaining slow-down seems to be coming from the add_vecs function, which sums together subarrays of two larger matrices. I've tried devectorizing and using BLAS functions, but haven't been able to produce better performance.
In any event, the improved code for dynam_serial is below:
function add_vecs(r::Array{Float32},h::Int,j::Int,E::Array{Float32},exp_v::Array{Float32},beta::Float32)
#views r=E[:,h,j] + beta*exp_v[:,j]
return r
end
function dynam_serial(E::Array{Float32},gm::Int,gz::Int,beta::Float32,Zprob::Array{Float32})
v = Array{Float32}(gm,gz)
fill!(v,E[cld(gm,2),cld(gm,2),cld(gz,2)])
Tv = Array{Float32}(gm,gz)
r = Array{Float32}(gm)
# Set parameters for the loop
convcrit = 0.0001 # chosen convergence criterion
diff = 1 # arbitrary initial value greater than convcrit
while diff>convcrit
exp_v=v*Zprob'
for h=1:gm
for j=1:gz
#views Tv[h,j]=findmax(add_vecs(r,h,j,E,exp_v,beta))[1]
end
end
diff = maximum(abs,Tv - v)
v=copy(Tv)
end
return Tv
end
If add_vecs seems to be the critical function, writing an explicit for loop could offer more optimization. How does the following benchmark:
function add_vecs!(r::Array{Float32},h::Int,j::Int,E::Array{Float32},
exp_v::Array{Float32},beta::Float32)
#inbounds for i=1:size(E,1)
r[i]=E[i,h,j] + beta*exp_v[i,j]
end
return r
end
UPDATE
To continue optimizing dynam_serial I have tried to remove more allocations. The result is:
function add_vecs_and_max!(gm::Int,r::Array{Float64},h::Int,j::Int,E::Array{Float64},
exp_v::Array{Float64},beta::Float64)
#inbounds for i=1:gm
r[i] = E[i,h,j]+beta*exp_v[i,j]
end
return findmax(r)[1]
end
function dynam_serial(E::Array{Float64},gm::Int,gz::Int,
beta::Float64,Zprob::Array{Float64})
v = Array{Float64}(gm,gz)
fill!(v,E[cld(gm,2),cld(gm,2),cld(gz,2)])
r = Array{Float64}(gm)
exp_v = Array{Float64}(gm,gz)
# Set parameters for the loop
convcrit = 0.0001 # chosen convergence criterion
diff = 1.0 # arbitrary initial value greater than convcrit
while diff>convcrit
A_mul_Bt!(exp_v,v,Zprob)
diff = -Inf
for h=1:gm
for j=1:gz
oldv = v[h,j]
newv = add_vecs_and_max!(gm,r,h,j,E,exp_v,beta)
v[h,j]= newv
diff = max(diff, oldv-newv, newv-oldv)
end
end
end
return v
end
Switching the functions to use Float64 should increase speed (as CPUs are inherently optimized for 64-bit word lengths). Also, using the mutating A_mul_Bt! directly saves another allocation. Avoiding the copy(...) by switching the arrays v and Tv.
How do these optimizations improve your running time?
2nd UPDATE
Updated the code in the UPDATE section to use findmax. Also, changed dynam_serial to use v without Tv, as there was no need to save the old version except for the diff calculation, which is now done inside the loop.
Here's the code I copied-and-pasted, provided by Dan Getz above. I include the array and scalar definitions exactly as I ran them. Performance was: 39.507005 seconds (11 allocations: 486.891 KiB) when running #time dynam_serial(E,gm,gz,beta,Zprob).
using SpecialFunctions
est_params = [.788,.288,.0034,.1519,.1615,.0041,.0077,.2,0.005,.7196]
r = 0.015
tau = 0.35
rho =est_params[1]
sigma =est_params[2]
delta = 0.15
gamma =est_params[3]
a_capital =est_params[4]
lambda1 =est_params[5]
lambda2 =est_params[6]
s =est_params[7]
theta =est_params[8]
mu =est_params[9]
p_bar_k_ss =est_params[10]
beta = (1+r)^(-1)
sigma_range = 4
gz = 19 #15 #19
gp = 29 #19 #29
gk = 37 #25 #37
lnz=collect(linspace(-sigma_range*sigma,sigma_range*sigma,gz))
z=exp.(lnz)
gk_m = fld(gk,2)
# Need to add mu somewhere to k_ss
k_ss = (theta*(1-tau)/(r+delta))^(1/(1-theta))
k=cat(1,map(i->k_ss*((1-delta)^i),collect(1:gk_m)),map(i->k_ss/((1-delta)^i),collect(1:gk_m)))
insert!(k,gk_m+1,k_ss)
sort!(k)
p_bar=p_bar_k_ss*k_ss
p = collect(linspace(-p_bar/2,p_bar,gp))
#Tauchen
N = length(z)
Z = zeros(N,1)
Zprob = zeros(Float64,N,N)
Z[N] = lnz[length(z)]
Z[1] = lnz[1]
zstep = (Z[N] - Z[1]) / (N - 1)
for i=2:(N-1)
Z[i] = Z[1] + zstep * (i - 1)
end
for a = 1 : N
for b = 1 : N
if b == 1
Zprob[a,b] = 0.5*erfc(-((Z[1] - mu - rho * Z[a] + zstep / 2) / sigma)/sqrt(2))
elseif b == N
Zprob[a,b] = 1 - 0.5*erfc(-((Z[N] - mu - rho * Z[a] - zstep / 2) / sigma)/sqrt(2))
else
Zprob[a,b] = 0.5*erfc(-((Z[b] - mu - rho * Z[a] + zstep / 2) / sigma)/sqrt(2)) -
0.5*erfc(-((Z[b] - mu - rho * Z[a] - zstep / 2) / sigma)/sqrt(2))
end
end
end
# Collecting tauchen results in a 2 element array of linspace and array; [2] gets array
# Zprob=collect(tauchen(gz, rho, sigma, mu, sigma_range))[2]
Zcumprob=zeros(Float64,gz,gz)
# 2 in cumsum! denotes the 2nd dimension, i.e. columns
cumsum!(Zcumprob, Zprob,2)
gm = gk * gp
control=zeros(gm,2)
for i=1:gk
control[(1+gp*(i-1)):(gp*i),1]=fill(k[i],(gp,1))
control[(1+gp*(i-1)):(gp*i),2]=p
end
endog=copy(control)
E=Array(Float64,gm,gm,gz)
for h=1:gm
for m=1:gm
for j=1:gz
# set the nonzero net debt indicator
if endog[h,2]<0
p_ind=1
else
p_ind=0
end
# set the investment indicator
if (control[m,1]-(1-delta)*endog[h,1])!=0
i_ind=1
else
i_ind=0
end
E[m,h,j] = (1-tau)*z[j]*(endog[h,1]^theta) + control[m,2]-endog[h,2]*(1+r*(1-tau)) +
delta*endog[h,1]*tau-(control[m,1]-(1-delta)*endog[h,1]) -
(i_ind*gamma*endog[h,1]+endog[h,1]*(a_capital/2)*(((control[m,1]-(1-delta)*endog[h,1])/endog[h,1])^2)) +
s*endog[h,2]*p_ind
elem = E[m,h,j]
if E[m,h,j]<0
E[m,h,j]=elem+lambda1*elem-.5*lambda2*elem^2
else
E[m,h,j]=elem
end
end
end
end
function add_vecs_and_max!(gm::Int,r::Array{Float64},h::Int,j::Int,E::Array{Float64},
exp_v::Array{Float64},beta::Float64)
maxr = -Inf
#inbounds for i=1:gm r[i] = E[i,h,j]+beta*exp_v[i,j]
maxr = max(r[i],maxr)
end
return maxr
end
function dynam_serial(E::Array{Float64},gm::Int,gz::Int,
beta::Float64,Zprob::Array{Float64})
v = Array{Float64}(gm,gz)
fill!(v,E[cld(gm,2),cld(gm,2),cld(gz,2)])
Tv = Array{Float64}(gm,gz)
r = Array{Float64}(gm)
exp_v = Array{Float64}(gm,gz)
# Set parameters for the loop
convcrit = 0.0001 # chosen convergence criterion
diff = 1.0 # arbitrary initial value greater than convcrit
while diff>convcrit
A_mul_Bt!(exp_v,v,Zprob)
diff = -Inf
for h=1:gm
for j=1:gz
Tv[h,j]=add_vecs_and_max!(gm,r,h,j,E,exp_v,beta)
diff = max(abs(Tv[h,j]-v[h,j]),diff)
end
end
(v,Tv)=(Tv,v)
end
return v
end
Now, here's another version of the algorithm and inputs. The functions are similar to what Dan Getz suggested, except that I use findmax rather than an iterated max function to find the array maximum. In the input construction, I am using both Float32 and mixing different bit-types together. However, I've consistently achieved better performance this way: 24.905569 seconds (1.81 k allocations: 46.829 MiB, 0.01% gc time). But it's not clear at all why.
using SpecialFunctions
est_params = [.788,.288,.0034,.1519,.1615,.0041,.0077,.2,0.005,.7196]
r = 0.015
tau = 0.35
rho =est_params[1]
sigma =est_params[2]
delta = 0.15
gamma =est_params[3]
a_capital =est_params[4]
lambda1 =est_params[5]
lambda2 =est_params[6]
s =est_params[7]
theta =est_params[8]
mu =est_params[9]
p_bar_k_ss =est_params[10]
beta = Float32((1+r)^(-1))
sigma_range = 4
gz = 19
gp = 29
gk = 37
lnz=collect(linspace(-sigma_range*sigma,sigma_range*sigma,gz))
z=exp(lnz)
gk_m = fld(gk,2)
# Need to add mu somewhere to k_ss
k_ss = (theta*(1-tau)/(r+delta))^(1/(1-theta))
k=cat(1,map(i->k_ss*((1-delta)^i),collect(1:gk_m)),map(i->k_ss/((1-delta)^i),collect(1:gk_m)))
insert!(k,gk_m+1,k_ss)
sort!(k)
p_bar=p_bar_k_ss*k_ss
p = collect(linspace(-p_bar/2,p_bar,gp))
#Tauchen
N = length(z)
Z = zeros(N,1)
Zprob = zeros(Float32,N,N)
Z[N] = lnz[length(z)]
Z[1] = lnz[1]
zstep = (Z[N] - Z[1]) / (N - 1)
for i=2:(N-1)
Z[i] = Z[1] + zstep * (i - 1)
end
for a = 1 : N
for b = 1 : N
if b == 1
Zprob[a,b] = 0.5*erfc(-((Z[1] - mu - rho * Z[a] + zstep / 2) / sigma)/sqrt(2))
elseif b == N
Zprob[a,b] = 1 - 0.5*erfc(-((Z[N] - mu - rho * Z[a] - zstep / 2) / sigma)/sqrt(2))
else
Zprob[a,b] = 0.5*erfc(-((Z[b] - mu - rho * Z[a] + zstep / 2) / sigma)/sqrt(2)) -
0.5*erfc(-((Z[b] - mu - rho * Z[a] - zstep / 2) / sigma)/sqrt(2))
end
end
end
# Collecting tauchen results in a 2 element array of linspace and array; [2] gets array
# Zprob=collect(tauchen(gz, rho, sigma, mu, sigma_range))[2]
Zcumprob=zeros(Float32,gz,gz)
# 2 in cumsum! denotes the 2nd dimension, i.e. columns
cumsum!(Zcumprob, Zprob,2)
gm = gk * gp
control=zeros(gm,2)
for i=1:gk
control[(1+gp*(i-1)):(gp*i),1]=fill(k[i],(gp,1))
control[(1+gp*(i-1)):(gp*i),2]=p
end
endog=copy(control)
E=Array(Float32,gm,gm,gz)
for h=1:gm
for m=1:gm
for j=1:gz
# set the nonzero net debt indicator
if endog[h,2]<0
p_ind=1
else
p_ind=0
end
# set the investment indicator
if (control[m,1]-(1-delta)*endog[h,1])!=0
i_ind=1
else
i_ind=0
end
E[m,h,j] = (1-tau)*z[j]*(endog[h,1]^theta) + control[m,2]-endog[h,2]*(1+r*(1-tau)) +
delta*endog[h,1]*tau-(control[m,1]-(1-delta)*endog[h,1]) -
(i_ind*gamma*endog[h,1]+endog[h,1]*(a_capital/2)*(((control[m,1]-(1-delta)*endog[h,1])/endog[h,1])^2)) +
s*endog[h,2]*p_ind
elem = E[m,h,j]
if E[m,h,j]<0
E[m,h,j]=elem+lambda1*elem-.5*lambda2*elem^2
else
E[m,h,j]=elem
end
end
end
end
function add_vecs!(gm::Int,r::Array{Float32},h::Int,j::Int,E::Array{Float32},
exp_v::Array{Float32},beta::Float32)
#inbounds #views for i=1:gm
r[i]=E[i,h,j] + beta*exp_v[i,j]
end
return r
end
function dynam_serial(E::Array{Float32},gm::Int,gz::Int,beta::Float32,Zprob::Array{Float32})
v = Array{Float32}(gm,gz)
fill!(v,E[cld(gm,2),cld(gm,2),cld(gz,2)])
Tv = Array{Float32}(gm,gz)
# Set parameters for the loop
convcrit = 0.0001 # chosen convergence criterion
diff = 1.00000 # arbitrary initial value greater than convcrit
iter=0
exp_v=Array{Float32}(gm,gz)
r=Array{Float32}(gm)
while diff>convcrit
A_mul_Bt!(exp_v,v,Zprob)
for h=1:gm
for j=1:gz
Tv[h,j]=findmax(add_vecs!(gm,r,h,j,E,exp_v,beta))[1]
end
end
diff = maximum(abs,Tv - v)
(v,Tv)=(Tv,v)
end
return v
end
I'm working on a function with three nested for loops that is way too slow for its intended use. The bottleneck is clearly the looping part - almost 100 % of the execution time is spent in the innermost loop.
The function takes a 2d matrix called rM as input and returns a 3d matrix called ec:
rows = size(rM, 1);
cols = size(rM, 2);
%preallocate.
ec = zeros(rows+1, cols, numRiskLevels);
ec(1, :, :) = 100;
for risk = minRisk:stepRisk:maxRisk;
for c = 1:cols,
for r = 2:rows+1,
ec(r, c, risk) = ec(r-1, c, risk) * (1 + risk * rM(r-1, c));
end
end
end
Any help on speeding up the for loops would be appreciated...
The problem is, that the inner loop is slowest, while it is also near-impossible to vectorize. As every iteration directly depends on the previous one.
The outer two are possible:
clc;
rM = rand(50);
rows = size(rM, 1);
cols = size(rM, 2);
minRisk = 1;
stepRisk = 1;
maxRisk = 100;
numRiskLevels = maxRisk/stepRisk;
%preallocate.
ec = zeros(rows+1, cols, numRiskLevels);
ec(1, :, :) = 100;
riskArray = (minRisk:stepRisk:maxRisk)';
tic
for r = 2:rows+1
tmp = riskArray * rM(r-1, :);
tmp = permute(tmp, [3 2 1]);
ec(r, :, :) = ec(r-1, :, :) .* (1 + tmp);
end
toc
%preallocate.
ec2 = zeros(rows+1, cols, numRiskLevels);
ec2(1, :, :) = 100;
tic
for risk = minRisk:stepRisk:maxRisk;
for c = 1:cols
for r = 2:rows+1
ec2(r, c, risk) = ec2(r-1, c, risk) * (1 + risk * rM(r-1, c));
end
end
end
toc
all(all(all(ec == ec2)))
But to my surprise, the vectorized code is indeed slower. (But maybe someone can improve the code, so I figured I leave it her for you.)
I have just tried to vectorize the outer loop, and actually noticed a significant speed increase. Of course it is hard to judge the speed of a script without knowing (the size of) the inputs but I would say this is a good starting point:
% Here you can change the input parameters
riskVec = 1:3:120;
rM = rand(50);
%preallocate and calculate non vectorized solution
ec2 = zeros(size(rM,2)+1, size(rM,1), max(riskVec));
ec2(1, :, :) = 100;
tic
for risk = riskVec
for c = 1:size(rM,2)
for r = 2:size(rM,1)+1
ec2(r, c, risk) = ec2(r-1, c, risk) * (1 + risk * rM(r-1, c));
end
end
end
t1=toc;
%preallocate and calculate vectorized solution
ec = zeros(size(rM,2)+1, size(rM,1), max(riskVec));
ec(1, :, :) = 100;
tic
for c = 1:size(rM,2)
for r = 2:size(rM,1)+1
ec(r, c, riskVec) = ec(r-1, c, riskVec) .* reshape(1 + riskVec * rM(r-1, c),[1 1 length(riskVec)]);
end
end
t2=toc;
% Check whether the vectorization is done correctly and show the timing results
if ec(:) == ec2(:)
t1
t2
end
The given output is:
t1 =
0.1288
t2 =
0.0408
So for this riskVec and rM it is about 3 times as fast as the non-vectorized solution.
I have written my own SHA1 implementation in MATLAB, and it gives correct hashes. However, it's very slow (a string a 1000 a's takes 9.9 seconds on my Core i7-2760QM), and I think the slowness is a result of how MATLAB implements bitwise logical operations (bitand, bitor, bitxor, bitcmp) and bitwise shifts (bitshift, bitrol, bitror) of integers.
Especially I wonder the need to construct fixed-point numeric objects for bitrol and bitror using fi command, because anyway in Intel x86 assembly there's rol and ror both for registers and memory addresses of all sizes. However, bitshift is quite fast (it doesn't need any fixed-point numeric costructs, a regular uint64 variable works fine), which makes the situation stranger: why in MATLAB bitrol and bitror need fixed-point numeric objects constructed with fi, whereas bitshift does not, when in assembly level it all comes down to shl, shr, rol and ror?
So, before writing this function in C/C++ as a .mex file, I'd be happy to know if there is any way to improve the performance of this function. I know there are some specific optimizations for SHA1, but that's not the issue, if the very basic implementation of bitwise rotations is so slow.
Testing a little bit with tic and toc, it's evident that what makes it slow are the loops in with bitrol and fi. There are two such loops:
%# Define some variables.
FFFFFFFF = uint64(hex2dec('FFFFFFFF'));
%# constants: K(1), K(2), K(3), K(4).
K(1) = uint64(hex2dec('5A827999'));
K(2) = uint64(hex2dec('6ED9EBA1'));
K(3) = uint64(hex2dec('8F1BBCDC'));
K(4) = uint64(hex2dec('CA62C1D6'));
W = uint64(zeros(1, 80));
... some other code here ...
%# First slow loop begins here.
for index = 17:80
W(index) = uint64(bitrol(fi(bitxor(bitxor(bitxor(W(index-3), W(index-8)), W(index-14)), W(index-16)), 0, 32, 0), 1));
end
%# First slow loop ends here.
H = sha1_handle_block_struct.H;
A = H(1);
B = H(2);
C = H(3);
D = H(4);
E = H(5);
%# Second slow loop begins here.
for index = 1:80
rotatedA = uint64(bitrol(fi(A, 0, 32, 0), 5));
if (index <= 20)
% alternative #1.
xorPart = bitxor(D, (bitand(B, (bitxor(C, D)))));
xorPart = bitand(xorPart, FFFFFFFF);
temp = rotatedA + xorPart + E + W(index) + K(1);
elseif ((index >= 21) && (index <= 40))
% FIPS.
xorPart = bitxor(bitxor(B, C), D);
xorPart = bitand(xorPart, FFFFFFFF);
temp = rotatedA + xorPart + E + W(index) + K(2);
elseif ((index >= 41) && (index <= 60))
% alternative #2.
xorPart = bitor(bitand(B, C), bitand(D, bitxor(B, C)));
xorPart = bitand(xorPart, FFFFFFFF);
temp = rotatedA + xorPart + E + W(index) + K(3);
elseif ((index >= 61) && (index <= 80))
% FIPS.
xorPart = bitxor(bitxor(B, C), D);
xorPart = bitand(xorPart, FFFFFFFF);
temp = rotatedA + xorPart + E + W(index) + K(4);
else
error('error in the code of sha1_handle_block.m!');
end
temp = bitand(temp, FFFFFFFF);
E = D;
D = C;
C = uint64(bitrol(fi(B, 0, 32, 0), 30));
B = A;
A = temp;
end
%# Second slow loop ends here.
Measuring with tic and toc, the entire computation of SHA1 hash of message abc takes on my laptop around 0.63 seconds, of which around 0.23 seconds is passed in the first slow loop and around 0.38 seconds in the second slow loop. So is there some way to optimize those loops in MATLAB before writing a .mex file?
There's this DataHash from the MATLAB File Exchange that calculates SHA-1 hashes lightning fast.
I ran the following code:
x = 'The quick brown fox jumped over the lazy dog'; %# Just a short sentence
y = repmat('a', [1, 1e6]); %# A million a's
opt = struct('Method', 'SHA-1', 'Format', 'HEX', 'Input', 'bin');
tic, x_hashed = DataHash(uint8(x), opt), toc
tic, y_hashed = DataHash(uint8(y), opt), toc
and got the following results:
x_hashed = F6513640F3045E9768B239785625CAA6A2588842
Elapsed time is 0.029250 seconds.
y_hashed = 34AA973CD4C4DAA4F61EEB2BDBAD27316534016F
Elapsed time is 0.020595 seconds.
I verified the results with a random online SHA-1 tool, and the calculation was indeed correct. Also, the 106 a's were hashed ~1.5 times faster than the first sentence.
So how does DataHash do it so fast??? Using the java.security.MessageDigest library, no less!
If you're interested with a fast MATLAB-friendly SHA-1 function, this is the way to go.
However, if this is just an exercise for implementing fast bit-level operations, then MATLAB doesn't really handle them efficiently, and in most cases you'll have to resort to MEX.
why in MATLAB bitrol and bitror need fixed-point numeric objects constructed with fi, whereas bitshift does not
bitrol and bitror are not part of the set of bitwise logic functions that are applicable for uints. They are part of the fixed-point toolbox, which also contains variants of bitand, bitshift etc that apply to fixed-point inputs.
A bitrol could be expressed as two bitshifts, a bitand and a bitor if you want to try using only the uint-functions. That might be even slower though.
As most MATLAB functions, bitand, bitor, bitxor are vectorized. So you get a lot faster if you give these function vector input rather than calling them in a loop over each element
Example:
%# create two sets of 10k random numbers
num = 10000;
hex = '0123456789ABCDEF';
A = uint64(hex2dec( hex(randi(16, [num 16])) ));
B = uint64(hex2dec( hex(randi(16, [num 16])) ));
%# compare loop vs. vectorized call
tic
C1 = zeros(size(A), class(A));
for i=1:numel(A)
C1(i) = bitxor(A(i),B(i));
end
toc
tic
C2 = bitxor(A,B);
toc
assert(isequal(C1,C2))
The timing was:
Elapsed time is 0.139034 seconds.
Elapsed time is 0.000960 seconds.
That's an order of magnitude faster!
The problem is, and as far as I can tell, the SHA-1 computation cannot be well vectorized. So you might not be able to take advantage of such vectorization.
As an experiment, I implemented a pure MATLAB-based funciton to compute such bit operations:
function num = my_bitops(op,A,B)
%# operation to perform: not, and, or, xor
if ischar(op)
op = str2func(op);
end
%# integer class: uint8, uint16, uint32, uint64
clss = class(A);
depth = str2double(clss(5:end));
%# bit exponents
e = 2.^(depth-1:-1:0);
%# convert to binary
b1 = logical(dec2bin(A,depth)-'0');
if nargin == 3
b2 = logical(dec2bin(B,depth)-'0');
end
%# perform binary operation
if nargin < 3
num = op(b1);
else
num = op(b1,b2);
end
%# convert back to integer
num = sum(bsxfun(#times, cast(num,clss), cast(e,clss)), 2, 'native');
end
Unfortunately, this was even worse in terms of performance:
tic, C1 = bitxor(A,B); toc
tic, C2 = my_bitops('xor',A,B); toc
assert(isequal(C1,C2))
The timing was:
Elapsed time is 0.000984 seconds.
Elapsed time is 0.485692 seconds.
Conclusion: write a MEX function or search the File Exchange to see if someone already did :)