stepan
/
dynare
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Adds a block Kalman filter using GPU

time-shift
Ferhat Mihoubi 2013-03-22 15:47:57 +01:00
parent ac6326758a
commit 2a51248832
1 changed files with 156 additions and 25 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

181
mex/sources/block_kalman_filter/block_kalman_filter.cc
View File

@ -26,10 +26,13 @@
#endif
#include "block_kalman_filter.h"
using namespace std;
//#define BLAS
#define DIRECT
#define BLAS
//#define CUBLAS
#ifdef CUBLAS
#include <cuda_runtime.h>
#include <cublas_v2.h>
#endif
void
mexDisp(mxArray* P)
{
@ -157,7 +160,7 @@ BlockKalmanFilter::BlockKalmanFilter(int nlhs, mxArray *plhs[], int nrhs, const
if (missing_observations)
{
if (! mxIsCell (prhs[0]))
DYN_MEX_FUNC_ERR_MSG_TXT("the first input argument of block_missing_observations_kalman_filter must be a Call Array.");
DYN_MEX_FUNC_ERR_MSG_TXT("the first input argument of block_missing_observations_kalman_filter must be a Cell Array.");
pdata_index = prhs[0];
if (! mxIsDouble (prhs[1]))
DYN_MEX_FUNC_ERR_MSG_TXT("the second input argument of block_missing_observations_kalman_filter must be a scalar.");
@ -234,14 +237,13 @@ BlockKalmanFilter::BlockKalmanFilter(int nlhs, mxArray *plhs[], int nrhs, const
*a = mxGetPr(pa);
tmp_a = (double*)mxMalloc(n * sizeof(double));
dF = 0.0; // det(F).
p_tmp = mxCreateDoubleMatrix(n, n_state, mxREAL);
*tmp = mxGetPr(p_tmp);
p_tmp1 = mxCreateDoubleMatrix(n, n_shocks, mxREAL);
tmp1 = mxGetPr(p_tmp1);
t = 0; // Initialization of the time index.
plik = mxCreateDoubleMatrix(smpl, 1, mxREAL);
lik = mxGetPr(plik);
Inf = mxGetInf();
Inf = mxGetInf();
LIK = 0.0; // Default value of the log likelihood.
notsteady = true; // Steady state flag.
F_singular = true;
@ -287,6 +289,22 @@ BlockKalmanFilter::BlockKalmanFilter(int nlhs, mxArray *plhs[], int nrhs, const
iw = (lapack_int*)mxMalloc(pp * sizeof(lapack_int));
ipiv = (lapack_int*)mxMalloc(pp * sizeof(lapack_int));
info = 0;
#ifdef BLAS || CUBLAS
p_tmp = mxCreateDoubleMatrix(n, n, mxREAL);
*tmp = mxGetPr(p_tmp);
p_P_t_t1 = mxCreateDoubleMatrix(n, n, mxREAL);
*P_t_t1 = mxGetPr(p_P_t_t1);
pK = mxCreateDoubleMatrix(n, n, mxREAL);
*K = mxGetPr(pK);
p_K_P = mxCreateDoubleMatrix(n, n, mxREAL);
*K_P = mxGetPr(p_K_P);
oldK = (double*)mxMalloc(n * n * sizeof(double));
*P_mf = (double*)mxMalloc(n * n * sizeof(double));
for (int i = 0; i < n * n; i++)
oldK[i] = Inf;
#else
p_tmp = mxCreateDoubleMatrix(n, n_state, mxREAL);
*tmp = mxGetPr(p_tmp);
p_P_t_t1 = mxCreateDoubleMatrix(n_state, n_state, mxREAL);
*P_t_t1 = mxGetPr(p_P_t_t1);
pK = mxCreateDoubleMatrix(n, pp, mxREAL);
@ -297,6 +315,7 @@ BlockKalmanFilter::BlockKalmanFilter(int nlhs, mxArray *plhs[], int nrhs, const
*P_mf = (double*)mxMalloc(n * pp * sizeof(double));
for (int i = 0; i < n * pp; i++)
oldK[i] = Inf;
#endif
}
void
@ -424,17 +443,17 @@ BlockKalmanFilter::block_kalman_filter(int nlhs, mxArray *plhs[], double *P_mf,
}
/* Computes the norm of iF */
double anorm = dlange("1", &size_d_index, &size_d_index, iF, &size_d_index, w);
/* Computes the norm of iF */
double anorm = dlange("1", &size_d_index, &size_d_index, iF, &size_d_index, w);
//mexPrintf("anorm = %f\n",anorm);
/* Modifies F in place with a LU decomposition */
dgetrf(&size_d_index, &size_d_index, iF, &size_d_index, ipiv, &info);
if (info != 0) fprintf(stderr, "dgetrf failure with error %d\n", (int) info);
/* Modifies F in place with a LU decomposition */
dgetrf(&size_d_index, &size_d_index, iF, &size_d_index, ipiv, &info);
if (info != 0) mexPrintf("dgetrf failure with error %d\n", (int) info);
/* Computes the reciprocal norm */
dgecon("1", &size_d_index, iF, &size_d_index, &anorm, &rcond, w, iw, &info);
if (info != 0) fprintf(stderr, "dgecon failure with error %d\n", (int) info);
/* Computes the reciprocal norm */
dgecon("1", &size_d_index, iF, &size_d_index, &anorm, &rcond, w, iw, &info);
if (info != 0) mexPrintf("dgecon failure with error %d\n", (int) info);
if (rcond < kalman_tol)
if (not_all_abs_F_bellow_crit(F, size_d_index * size_d_index, kalman_tol)) //~all(abs(F(:))<kalman_tol)
@ -506,7 +525,7 @@ BlockKalmanFilter::block_kalman_filter(int nlhs, mxArray *plhs[], double *P_mf,
//iF = inv(F);
//int lwork = 4/*2*/* pp;
dgetri(&size_d_index, iF, &size_d_index, ipiv, w, &lw, &info);
if (info != 0) fprintf(stderr, "dgetri failure with error %d\n", (int) info);
if (info != 0) mexPrintf("dgetri failure with error %d\n", (int) info);
//lik(t) = log(dF)+transpose(v)*iF*v;
#ifdef USE_OMP
@ -628,14 +647,126 @@ BlockKalmanFilter::block_kalman_filter(int nlhs, mxArray *plhs[], double *P_mf,
double one = 1.0;
double zero = 0.0;
memcpy(P, QQ, n * n *sizeof(double));
dsymm("R", "U", &n, &n,
&one, P_t_t1, &n,
T, &n, &zero,
tmp, &n);
dgemm("N", "T", &n, &n,
&n, &one, tmp, &n,
T, &n, &one,
P, &n);
blas_int n_b = n;
/*mexPrintf("sizeof(n_b)=%d, n_b=%d, sizeof(n)=%d, n=%d\n",sizeof(n_b),n_b,sizeof(n),n);
mexEvalString("drawnow;");*/
dsymm("R", "U", &n_b, &n_b,
&one, P_t_t1, &n_b,
T, &n_b, &zero,
tmp, &n_b);
dgemm("N", "T", &n_b, &n_b,
&n_b, &one, tmp, &n_b,
T, &n_b, &one,
P, &n_b);
#else
#ifdef CUBLAS
for (int i = 0; i < n; i++)
for (int j = i; j < n; j++)
{
double res = 0.0;
//int j_pp = j * pp;
for (int k = 0; k < size_d_index; k++)
res += K[i + k * n] * P_mf[k + j * size_d_index];
K_P[i * n + j] = K_P[j * n + i] = res;
}
//#pragma omp parallel for shared(P, K_P, P_t_t1)
for (int i = size_d_index; i < n; i++)
for (int j = i; j < n; j++)
{
unsigned int k = i * n + j;
P_t_t1[j * n + i] = P_t_t1[k] = P[k] - K_P[k];
}
mexPrintf("CudaBLAS\n");
mexEvalString("drawnow;");
double one = 1.0;
double zero = 0.0;
cublasStatus_t status;
cublasHandle_t handle;
status = cublasCreate(&handle);
if (status != CUBLAS_STATUS_SUCCESS)
{
mexPrintf("!!!! CUBLAS initialization error\n");
return false;
}
/*int device;
cudaGetDevice(&device);*/
int n2 = n * n;
double* d_A = 0;
double* d_B = 0;
double* d_C = 0;
double* d_D = 0;
// Allocate device memory for the matrices
if (cudaMalloc((void**)&d_A, n2 * sizeof(double)) != cudaSuccess)
{
mexPrintf("!!!! device memory allocation error (allocate A)\n");
return false;
}
if (cudaMalloc((void**)&d_B, n2 * sizeof(d_B[0])) != cudaSuccess)
{
mexPrintf("!!!! device memory allocation error (allocate B)\n");
return false;
}
if (cudaMalloc((void**)&d_C, n2 * sizeof(d_C[0])) != cudaSuccess)
{
mexPrintf("!!!! device memory allocation error (allocate C)\n");
return false;
}
if (cudaMalloc((void**)&d_D, n2 * sizeof(d_D[0])) != cudaSuccess)
{
mexPrintf("!!!! device memory allocation error (allocate D)\n");
return false;
}
// Initialize the device matrices with the host matrices
status = cublasSetVector(n2, sizeof(P_t_t1[0]), P_t_t1, 1, d_A, 1);
if (status != CUBLAS_STATUS_SUCCESS)
{
mexPrintf("!!!! device access error (write A)\n");
return false;
}
status = cublasSetVector(n2, sizeof(T[0]), T, 1, d_B, 1);
if (status != CUBLAS_STATUS_SUCCESS)
{
mexPrintf("!!!! device access error (write B)\n");
return false;
}
status = cublasSetVector(n2, sizeof(tmp[0]), tmp, 1, d_C, 1);
if (status != CUBLAS_STATUS_SUCCESS)
{
mexPrintf("!!!! device access error (write C)\n");
return false;
}
mexPrintf("just before calling\n");
mexEvalString("drawnow;");
status = cublasSetVector(n2, sizeof(QQ[0]), QQ, 1, d_D, 1);
if (status != CUBLAS_STATUS_SUCCESS)
{
mexPrintf("!!!! device access error (write D)\n");
return false;
}
// Performs operation using plain C code
cublasDsymm(handle, CUBLAS_SIDE_RIGHT, CUBLAS_FILL_MODE_UPPER, n, n,
&one, d_A, n,
d_B, n, &zero,
d_C, n);
/*dgemm("N", "T", &n_b, &n_b,
&n_b, &one, tmp, &n_b,
T, &n_b, &one,
P, &n_b);*/
cublasDgemm(handle, CUBLAS_OP_N, CUBLAS_OP_T, n, n,
n, &one, d_C, n,
d_B, n, &one,
d_D, n);
//double_symm(n, &one, h_A, h_B, &zero, h_C);
status = cublasGetVector(n2, sizeof(P[0]), d_D, 1, P, 1);
if (status != CUBLAS_STATUS_SUCCESS)
{
mexPrintf("!!!! device access error (read P)\n");
return false;
}
#else
#ifdef USE_OMP
#pragma omp parallel for shared(K_P) num_threads(atoi(getenv("DYNARE_NUM_THREADS")))
@ -717,7 +848,7 @@ BlockKalmanFilter::block_kalman_filter(int nlhs, mxArray *plhs[], double *P_mf,
P[i + j * n] = P[j + i * n];
}
#endif
#endif
if (t >= no_more_missing_observations)
{
double max_abs = 0.0;