我一直在尝试使用Fortran中的ACC和OpenMP进行并行化。现在我正尝试在Matlab中实现相同的功能。我发现很有趣的是,使用GPU在Matlab中实现循环并行化似乎非常困难。显然,唯一的方法是使用arrayfun
函数。但我可能是错的。
从概念上讲,我想知道为什么在Matlab中使用GPU不比在Fortran中更简单。在更实际的层面上,我想知道如何在下面的简单代码中使用GPU。
下面,我分享了三个代码和基准测试:
- Fortran OpenMP 代码
- Fortran ACC 代码
- Matlab parfor 代码
- Matlab CUDA (?) 这是我不知道如何做的。
Fortran OpenMP 代码:
program rbc
use omp_lib ! For timing
use tools
implicit none
real, parameter :: beta = 0.984, eta = 2, alpha = 0.35, delta = 0.01, &
rho = 0.95, sigma = 0.005, zmin=-0.0480384, zmax=0.0480384;
integer, parameter :: nz = 4, nk=4800;
real :: zgrid(nz), kgrid(nk), t_tran_z(nz,nz), tran_z(nz,nz);
real :: kmax, kmin, tol, dif, c(nk), r(nk), w(nk);
real, dimension(nk,nz) :: v=0., v0=0., ev=0., c0=0.;
integer :: i, iz, ik, cnt;
logical :: ind(nk);
real(kind=8) :: start, finish ! For timing
real :: tmpmax, c1
call omp_set_num_threads(12)
!Grid for productivity z
! [1 x 4] grid of values for z
call linspace(zmin,zmax,nz,zgrid)
zgrid = exp(zgrid)
! [4 x 4] Markov transition matrix of z
tran_z(1,1) = 0.996757
tran_z(1,2) = 0.00324265
tran_z(1,3) = 0
tran_z(1,4) = 0
tran_z(2,1) = 0.000385933
tran_z(2,2) = 0.998441
tran_z(2,3) = 0.00117336
tran_z(2,4) = 0
tran_z(3,1) = 0
tran_z(3,2) = 0.00117336
tran_z(3,3) = 0.998441
tran_z(3,4) = 0.000385933
tran_z(4,1) = 0
tran_z(4,2) = 0
tran_z(4,3) = 0.00324265
tran_z(4,4) = 0.996757
! Grid for capital k
kmin = 0.95*(1/(alpha*zgrid(1)))*((1/beta)-1+delta)**(1/(alpha-1));
kmax = 1.05*(1/(alpha*zgrid(nz)))*((1/beta)-1+delta)**(1/(alpha-1));
! [1 x 4800] grid of possible values of k
call linspace(kmin, kmax, nk, kgrid)
! Compute initial wealth c0(k,z)
do iz=1,nz
c0(:,iz) = zgrid(iz)*kgrid**alpha + (1-delta)*kgrid;
end do
dif = 10000
tol = 1e-8
cnt = 1
do while(dif>tol)
!$omp parallel do default(shared) private(ik,iz,i,tmpmax,c1)
do ik=1,nk;
do iz = 1,nz;
tmpmax = -huge(0.)
do i = 1,nk
c1 = c0(ik,iz) - kgrid(i)
if(c1<0) exit
c1 = c1**(1-eta)/(1-eta)+ev(i,iz)
if(tmpmax<c1) tmpmax = c1
end do
v(ik,iz) = tmpmax
end do
end do
!$omp end parallel do
ev = beta*matmul(v,tran_z)
dif = maxval(abs(v-v0))
v0 = v
if(mod(cnt,1)==0) write(*,*) cnt, ':', dif
cnt = cnt+1
end do
end program
Fortran ACC:
只需将上述代码中的主循环语法替换为:
do while(dif>tol)
!$acc kernels
!$acc loop gang
do ik=1,nk;
!$acc loop gang
do iz = 1,nz;
tmpmax = -huge(0.)
do i = 1,nk
c1 = c0(ik,iz) - kgrid(i)
if(c1<0) exit
c1 = c1**(1-eta)/(1-eta)+ev(i,iz)
if(tmpmax<c1) tmpmax = c1
end do
v(ik,iz) = tmpmax
end do
end do
!$acc end kernels
ev = beta*matmul(v,tran_z)
dif = maxval(abs(v-v0))
v0 = v
if(mod(cnt,1)==0) write(*,*) cnt, ':', dif
cnt = cnt+1
end do
Matlab parfor: (我知道下面的代码可以使用向量化语法使其更快,但整个练习的重点是比较循环速度。)
tic;
beta = 0.984;
eta = 2;
alpha = 0.35;
delta = 0.01;
rho = 0.95;
sigma = 0.005;
zmin=-0.0480384;
zmax=0.0480384;
nz = 4;
nk=4800;
v=zeros(nk,nz);
v0=zeros(nk,nz);
ev=zeros(nk,nz);
c0=zeros(nk,nz);
%Grid for productivity z
%[1 x 4] grid of values for z
zgrid = linspace(zmin,zmax,nz);
zgrid = exp(zgrid);
% [4 x 4] Markov transition matrix of z
tran_z(1,1) = 0.996757;
tran_z(1,2) = 0.00324265;
tran_z(1,3) = 0;
tran_z(1,4) = 0;
tran_z(2,1) = 0.000385933;
tran_z(2,2) = 0.998441;
tran_z(2,3) = 0.00117336;
tran_z(2,4) = 0;
tran_z(3,1) = 0;
tran_z(3,2) = 0.00117336;
tran_z(3,3) = 0.998441;
tran_z(3,4) = 0.000385933;
tran_z(4,1) = 0;
tran_z(4,2) = 0;
tran_z(4,3) = 0.00324265;
tran_z(4,4) = 0.996757;
% Grid for capital k
kmin = 0.95*(1/(alpha*zgrid(1)))*((1/beta)-1+delta)^(1/(alpha-1));
kmax = 1.05*(1/(alpha*zgrid(nz)))*((1/beta)-1+delta)^(1/(alpha-1));
% [1 x 4800] grid of possible values of k
kgrid = linspace(kmin, kmax, nk);
% Compute initial wealth c0(k,z)
for iz=1:nz
c0(:,iz) = zgrid(iz)*kgrid.^alpha + (1-delta)*kgrid;
end
dif = 10000;
tol = 1e-8;
cnt = 1;
while dif>tol
parfor ik=1:nk
for iz = 1:nz
tmpmax = -intmax;
for i = 1:nk
c1 = c0(ik,iz) - kgrid(i);
if (c1<0)
continue
end
c1 = c1^(1-eta)/(1-eta)+ev(i,iz);
if tmpmax<c1
tmpmax = c1;
end
end
v(ik,iz) = tmpmax;
end
end
ev = beta*v*tran_z;
dif = max(max(abs(v-v0)));
v0 = v;
if mod(cnt,1)==0
fprintf('%1.5f : %1.5f \n', [cnt dif])
end
cnt = cnt+1;
end
toc
Matlab CUDA:
这是我不知道如何编码的部分。使用arrayfun
是唯一的方法吗?在Fortran中,从OpenMP转换到OpenACC非常简单。在Matlab中,从parfor转换为GPU循环没有简单的方法吗?
代码之间的时间比较:
Fortran OpenMP: 83.1 seconds
Fortran ACC: 2.4 seconds
Matlab parfor: 1182 seconds
最后一点,我应该说上面的代码解决了一个简单的实际商业周期模型,并基于这个写成。
gpuArray()
将所有内容放入GPU,然后使用经典的矢量化语法。如果你不能以矢量化方式编写代码,那么arrayfun
是一种更繁琐的替代方案。 - Nicky Mattsson