//===================================================================================================================
//===================================================================================================================
//===================================================================================================================
extern "C" void
magma_dlarft_sm32x32_batched(magma_int_t n, magma_int_t k,
double **v_array, magma_int_t ldv,
double **tau_array,
double **T_array, magma_int_t ldt,
magma_int_t batchCount, magma_queue_t queue)
{
if ( k <= 0) return;
//==================================
// GEMV
//==================================
#define USE_GEMV2
#define use_gemm_larft_sm32
#if defined(use_gemm_larft_sm32)
magma_dgemm_batched( MagmaConjTrans, MagmaNoTrans,
k, k, n,
MAGMA_D_ONE, v_array, ldv,
v_array, ldv,
MAGMA_D_ZERO, T_array, ldt,
batchCount, queue );
magmablas_dlaset_batched( MagmaLower, k, k,
MAGMA_D_ZERO, MAGMA_D_ZERO,
T_array, ldt, batchCount, queue );
#else
#if 1
for (magma_int_t i=0; i < k; i++)
{
//W(1:i-1) := - tau(i) * V(i:n,1:i-1)' * V(i:n,i)
//T( i, i ) = tau( i )
//custom implementation.
#ifdef USE_GEMV2
magmablas_dlarft_gemvrowwise_batched( n-i, i,
tau_array,
v_array, ldv,
T_array, ldt,
batchCount, queue);
#else
magmablas_dlarft_gemvcolwise_batched( n-i, i, v_array, ldv, T_array, ldt, tau_array, batchCount, queue);
#endif
}
#else
//seems to be very slow when k=32 while the one by one loop above is faster
dlarft_gemv_loop_inside_kernel_batched(n, k, tau_array, v_array, ldv, T_array, ldt, batchCount, queue);
#endif
#endif
//==================================
// TRMV
//==================================
//T(1:i-1,i) := T(1:i-1,1:i-1) * W(1:i-1) i=[1:k]
magmablas_dlarft_dtrmv_sm32x32_batched(k, k, tau_array, T_array, ldt, T_array, ldt, batchCount, queue);
}
/***************************************************************************//**
Purpose
-------
DGEQRF computes a QR factorization of a real M-by-N matrix A:
A = Q * R.
Arguments
---------
@param[in]
m INTEGER
The number of rows of the matrix A. M >= 0.
@param[in]
n INTEGER
The number of columns of the matrix A. N >= 0.
@param[in,out]
dA_array Array of pointers, dimension (batchCount).
Each is a DOUBLE PRECISION array on the GPU, dimension (LDDA,N)
On entry, the M-by-N matrix A.
On exit, the elements on and above the diagonal of the array
contain the min(M,N)-by-N upper trapezoidal matrix R (R is
upper triangular if m >= n); the elements below the diagonal,
with the array TAU, represent the orthogonal matrix Q as a
product of min(m,n) elementary reflectors (see Further
Details).
@param[in]
ldda INTEGER
The leading dimension of the array dA. LDDA >= max(1,M).
To benefit from coalescent memory accesses LDDA must be
divisible by 16.
@param[in,out]
dR_array Array of pointers, dimension (batchCount).
Each is a DOUBLE PRECISION array on the GPU, dimension (LDDR, N/NB)
dR should be of size (LDDR, N) when provide_RT > 0 and
of size (LDDT, NB) otherwise. NB is the local blocking size.
On exit, the elements of R are stored in dR only when provide_RT > 0.
@param[in]
lddr INTEGER
The leading dimension of the array dR.
LDDR >= min(M,N) when provide_RT == 1
otherwise LDDR >= min(NB, min(M,N)).
NB is the local blocking size.
To benefit from coalescent memory accesses LDDR must be
divisible by 16.
@param[in,out]
dT_array Array of pointers, dimension (batchCount).
Each is a DOUBLE PRECISION array on the GPU, dimension (LDDT, N/NB)
dT should be of size (LDDT, N) when provide_RT > 0 and
of size (LDDT, NB) otherwise. NB is the local blocking size.
On exit, the elements of T are stored in dT only when provide_RT > 0.
@param[in]
lddt INTEGER
The leading dimension of the array dT.
LDDT >= min(NB,min(M,N)). NB is the local blocking size.
To benefit from coalescent memory accesses LDDR must be
divisible by 16.
@param[out]
dtau_array Array of pointers, dimension (batchCount).
Each is a DOUBLE PRECISION array, dimension (min(M,N))
The scalar factors of the elementary reflectors (see Further
Details).
@param[in]
provide_RT INTEGER
provide_RT = 0 no R and no T in output.
dR and dT are used as local workspace to store the R and T of each step.
provide_RT = 1 the whole R of size (min(M,N), N) and the nbxnb block of T are provided in output.
provide_RT = 2 the nbxnb diag block of R and of T are provided in output.
@param[out]
info_array Array of INTEGERs, dimension (batchCount), for corresponding matrices.
- = 0: successful exit
- < 0: if INFO = -i, the i-th argument had an illegal value
or another error occured, such as memory allocation failed.
@param[in]
batchCount INTEGER
The number of matrices to operate on.
@param[in]
queue magma_queue_t
Queue to execute in.
Further Details
---------------
The matrix Q is represented as a product of elementary reflectors
Q = H(1) H(2) . . . H(k), where k = min(m,n).
Each H(i) has the form
H(i) = I - tau * v * v'
where tau is a real scalar, and v is a real vector with
v(1:i-1) = 0 and v(i) = 1; v(i+1:m) is stored on exit in A(i+1:m,i),
and tau in TAU(i).
@ingroup magma_geqrf_batched
*******************************************************************************/
extern "C" magma_int_t
magma_dgeqrf_expert_batched(
magma_int_t m, magma_int_t n,
double **dA_array, magma_int_t ldda,
double **dR_array, magma_int_t lddr,
double **dT_array, magma_int_t lddt,
double **dtau_array, magma_int_t provide_RT,
magma_int_t *info_array, magma_int_t batchCount, magma_queue_t queue)
{
#define dA(i, j) (dA + (i) + (j)*ldda)
/* Local Parameter */
magma_int_t nb = magma_get_dgeqrf_batched_nb(m);
magma_int_t nnb = 8;
magma_int_t min_mn = min(m, n);
/* Check arguments */
cudaMemset(info_array, 0, batchCount*sizeof(magma_int_t));
magma_int_t arginfo = 0;
if (m < 0)
arginfo = -1;
else if (n < 0)
arginfo = -2;
else if (ldda < max(1,m))
arginfo = -4;
else if (lddr < min_mn && provide_RT == 1)
arginfo = -6;
else if (lddr < min(min_mn, nb))
arginfo = -6;
else if (lddt < min(min_mn, nb))
arginfo = -8;
if (arginfo != 0) {
magma_xerbla( __func__, -(arginfo) );
return arginfo;
}
/* Quick return if possible */
if (m == 0 || n == 0)
if (min_mn == 0 ) return arginfo;
if ( m > 2048 || n > 2048 ) {
printf("=========================================================================================\n");
printf(" WARNING batched routines are designed for small sizes it might be better to use the\n Native/Hybrid classical routines if you want performance\n");
printf("=========================================================================================\n");
}
magma_int_t i, k, ib=nb, jb=nnb, offset_RT=0, use_stream;
magma_int_t ldw, offset;
double **dW0_displ = NULL;
double **dW1_displ = NULL;
double **dW2_displ = NULL;
double **dW3_displ = NULL;
double **dW4_displ = NULL;
double **dW5_displ = NULL;
double **dR_displ = NULL;
double **dT_displ = NULL;
double *dwork = NULL;
double **cpuAarray = NULL;
double **cpuTarray = NULL;
magma_malloc((void**)&dW0_displ, batchCount * sizeof(*dW0_displ));
magma_malloc((void**)&dW1_displ, batchCount * sizeof(*dW1_displ));
magma_malloc((void**)&dW2_displ, batchCount * sizeof(*dW2_displ));
magma_malloc((void**)&dW3_displ, batchCount * sizeof(*dW3_displ));
magma_malloc((void**)&dW4_displ, batchCount * sizeof(*dW4_displ));
magma_malloc((void**)&dW5_displ, batchCount * sizeof(*dW5_displ));
magma_malloc((void**)&dR_displ, batchCount * sizeof(*dR_displ));
magma_malloc((void**)&dT_displ, batchCount * sizeof(*dT_displ));
magma_dmalloc(&dwork, (2 * nb * n) * batchCount);
magma_malloc_cpu((void**) &cpuAarray, batchCount*sizeof(double*));
magma_malloc_cpu((void**) &cpuTarray, batchCount*sizeof(double*));
/* check allocation */
if ( dW0_displ == NULL || dW1_displ == NULL || dW2_displ == NULL ||
dW3_displ == NULL || dW4_displ == NULL || dW5_displ == NULL ||
dR_displ == NULL || dT_displ == NULL || dwork == NULL ||
cpuAarray == NULL || cpuTarray == NULL ) {
magma_free(dW0_displ);
magma_free(dW1_displ);
magma_free(dW2_displ);
magma_free(dW3_displ);
magma_free(dW4_displ);
magma_free(dW5_displ);
magma_free(dR_displ);
magma_free(dT_displ);
magma_free(dwork);
magma_free_cpu(cpuAarray);
magma_free_cpu(cpuTarray);
magma_int_t info = MAGMA_ERR_DEVICE_ALLOC;
magma_xerbla( __func__, -(info) );
return info;
}
magma_ddisplace_pointers(dR_displ, dR_array, lddr, 0, 0, batchCount, queue);
magma_ddisplace_pointers(dT_displ, dT_array, lddt, 0, 0, batchCount, queue);
// set dwork to zero because our GEMM routine does propagate NAN when C=betaC+alphaA*B and beta=0
magmablas_dlaset_q( MagmaFull, 2*nb, n*batchCount, MAGMA_D_ZERO, MAGMA_D_ZERO, dwork, 2*nb, queue );
// set dR and dT to zero. if provide_RT == 0 only a tile of size nbxnb is used and overwritten at each step
magmablas_dlaset_batched( MagmaFull, lddr, (provide_RT > 0 ? n:min(min_mn,nb)), MAGMA_D_ZERO, MAGMA_D_ZERO, dR_displ, lddr, batchCount, queue );
magmablas_dlaset_batched( MagmaFull, lddt, (provide_RT > 0 ? n:min(min_mn,nb)), MAGMA_D_ZERO, MAGMA_D_ZERO, dT_displ, lddt, batchCount, queue );
/*
if ( provide_RT > 0 )
{
magmablas_dlaset_q( MagmaFull, lddr, n*batchCount, MAGMA_D_ZERO, MAGMA_D_ZERO, dR, lddr, queue );
magmablas_dlaset_q( MagmaFull, lddt, n*batchCount, MAGMA_D_ZERO, MAGMA_D_ZERO, dT, lddt, queue );
}
else
{
magmablas_dlaset_q( MagmaFull, lddr, nb*batchCount, MAGMA_D_ZERO, MAGMA_D_ZERO, dR, lddr, queue );
magmablas_dlaset_q( MagmaFull, lddt, nb*batchCount, MAGMA_D_ZERO, MAGMA_D_ZERO, dT, lddt, queue );
}
*/
magma_int_t streamid;
const magma_int_t nbstreams=10;
magma_queue_t queues[nbstreams];
for (i=0; i < nbstreams; i++) {
magma_device_t cdev;
magma_getdevice( &cdev );
magma_queue_create( cdev, &queues[i] );
}
magma_getvector( batchCount, sizeof(double*), dA_array, 1, cpuAarray, 1, queue);
magma_getvector( batchCount, sizeof(double*), dT_array, 1, cpuTarray, 1, queue);
for (i=0; i < min_mn; i += nb)
{
ib = min(nb, min_mn-i);
//===============================================
// panel factorization
//===============================================
magma_ddisplace_pointers(dW0_displ, dA_array, ldda, i, i, batchCount, queue);
magma_ddisplace_pointers(dW2_displ, dtau_array, 1, i, 0, batchCount, queue);
if ( provide_RT > 0 )
{
offset_RT = i;
magma_ddisplace_pointers(dR_displ, dR_array, lddr, (provide_RT == 1 ? offset_RT:0), offset_RT, batchCount, queue);
magma_ddisplace_pointers(dT_displ, dT_array, lddt, 0, offset_RT, batchCount, queue);
}
//dwork is used in panel factorization and trailing matrix update
//dW4_displ, dW5_displ are used as workspace and configured inside
magma_dgeqrf_panel_batched(m-i, ib, jb,
dW0_displ, ldda,
dW2_displ,
dT_displ, lddt,
dR_displ, lddr,
dW1_displ,
dW3_displ,
dwork,
dW4_displ, dW5_displ,
info_array,
batchCount, queue);
//===============================================
// end of panel
//===============================================
//===============================================
// update trailing matrix
//===============================================
if ( (n-ib-i) > 0)
{
//dwork is used in panel factorization and trailing matrix update
//reset dW4_displ
ldw = nb;
magma_dset_pointer( dW4_displ, dwork, 1, 0, 0, ldw*n, batchCount, queue );
offset = ldw*n*batchCount;
magma_dset_pointer( dW5_displ, dwork + offset, 1, 0, 0, ldw*n, batchCount, queue );
// set the diagonal of v as one and the upper triangular part as zero already set inside geqrf_panel
//magmablas_dlaset_batched( MagmaUpper, ib, ib, MAGMA_D_ZERO, MAGMA_D_ONE, dW0_displ, ldda, batchCount, queue );
//magma_ddisplace_pointers(dW2_displ, dtau_array, 1, i, 0, batchCount, queue);
// it is faster since it is using BLAS-3 GEMM routines, different from lapack implementation
magma_dlarft_batched(m-i, ib, 0,
dW0_displ, ldda,
dW2_displ,
dT_displ, lddt,
dW4_displ, nb*lddt,
batchCount, queue);
// perform C = (I-V T^H V^H) * C, C is the trailing matrix
//-------------------------------------------
// USE STREAM GEMM
//-------------------------------------------
use_stream = magma_drecommend_cublas_gemm_stream(MagmaNoTrans, MagmaNoTrans, m-i-ib, n-i-ib, ib);
if ( use_stream )
{
magma_queue_sync(queue);
for (k=0; k < batchCount; k++)
{
streamid = k%nbstreams;
// the queue gemm must take cpu pointer
magma_dlarfb_gpu_gemm( MagmaLeft, MagmaConjTrans, MagmaForward, MagmaColumnwise,
m-i, n-i-ib, ib,
cpuAarray[k] + i + i * ldda, ldda,
cpuTarray[k] + offset_RT*lddt, lddt,
cpuAarray[k] + i + (i+ib) * ldda, ldda,
dwork + nb * n * k, -1,
dwork + nb * n * batchCount + nb * n * k, -1, queues[streamid] );
}
// need to synchronise to be sure that panel does not start before
// finishing the update at least of the next panel
// if queue is NULL, no need to sync
if ( queue != NULL ) {
for (magma_int_t s=0; s < nbstreams; s++)
magma_queue_sync(queues[s]);
}
}
//-------------------------------------------
// USE BATCHED GEMM
//-------------------------------------------
else
{
//direct trailing matrix in dW1_displ
magma_ddisplace_pointers(dW1_displ, dA_array, ldda, i, i+ib, batchCount, queue);
magma_dlarfb_gemm_batched(
MagmaLeft, MagmaConjTrans, MagmaForward, MagmaColumnwise,
m-i, n-i-ib, ib,
(const double**)dW0_displ, ldda,
(const double**)dT_displ, lddt,
dW1_displ, ldda,
dW4_displ, ldw,
dW5_displ, ldw,
batchCount, queue );
}
}// update the trailing matrix
//===============================================
// copy dR back to V after the trailing matrix update,
// only when provide_RT=0 otherwise the nbxnb block of V is set to diag=1/0
// The upper portion of V could be set totaly to 0 here
if ( provide_RT == 0 )
{
magmablas_dlacpy_batched( MagmaUpper, ib, ib, dR_displ, lddr, dW0_displ, ldda, batchCount, queue );
}
}
magma_queue_sync(queue);
for (k=0; k < nbstreams; k++) {
magma_queue_destroy( queues[k] );
}
magma_free(dW0_displ);
magma_free(dW1_displ);
magma_free(dW2_displ);
magma_free(dW3_displ);
magma_free(dW4_displ);
magma_free(dW5_displ);
magma_free(dR_displ);
magma_free(dT_displ);
magma_free(dwork);
magma_free_cpu(cpuAarray);
magma_free_cpu(cpuTarray);
return arginfo;
}
//===================================================================================================================
//===================================================================================================================
//===================================================================================================================
extern "C" magma_int_t
magma_dlarft_batched(magma_int_t n, magma_int_t k, magma_int_t stair_T,
double **v_array, magma_int_t ldv,
double **tau_array, double **T_array, magma_int_t ldt,
double **work_array, magma_int_t lwork,
magma_int_t batchCount, magma_queue_t queue)
{
double c_one = MAGMA_D_ONE;
double c_zero = MAGMA_D_ZERO;
if ( k <= 0) return 0;
if ( stair_T > 0 && k <= stair_T) return 0;
magma_int_t maxnb = max_shared_bsiz;
if ( lwork < k*ldt)
{
magma_xerbla( __func__, -(10) );
return -10;
}
if ( stair_T > 0 && stair_T > maxnb)
{
magma_xerbla( __func__, -(3) );
return -3;
}
magma_int_t DEBUG=0;
magma_int_t nb = stair_T == 0 ? min(k,maxnb) : stair_T;
magma_int_t i, j, prev_n, mycol, rows;
double **dW1_displ = NULL;
double **dW2_displ = NULL;
double **dW3_displ = NULL;
double **dTstep_array = NULL;
magma_malloc((void**)&dW1_displ, batchCount * sizeof(*dW1_displ));
magma_malloc((void**)&dW2_displ, batchCount * sizeof(*dW2_displ));
magma_malloc((void**)&dW3_displ, batchCount * sizeof(*dW3_displ));
magma_malloc((void**)&dTstep_array, batchCount * sizeof(*dTstep_array));
//double *Tstep = k > nb ? work : T;
if (k > nb)
{
magma_ddisplace_pointers(dTstep_array, work_array, lwork, 0, 0, batchCount, queue);
}
else
{
magma_ddisplace_pointers(dTstep_array, T_array, ldt, 0, 0, batchCount, queue);
}
//magma_int_t ldtstep = k > nb ? k : ldt;
magma_int_t ldtstep = ldt; //a enlever
// stair_T = 0 meaning all T
// stair_T > 0 meaning the triangular portion of T has been computed.
// the value of stair_T is the nb of these triangulars
//GEMV compute the whole triangular upper portion of T (phase 1)
// TODO addcublas to check perf
magma_dgemm_batched( MagmaConjTrans, MagmaNoTrans,
k, k, n,
c_one, v_array, ldv,
v_array, ldv,
c_zero, dTstep_array, ldtstep,
batchCount, queue );
magmablas_dlaset_batched( MagmaLower, k, k, MAGMA_D_ZERO, MAGMA_D_ZERO, dTstep_array, ldtstep, batchCount, queue );
// no need for it as T is expected to be lower zero
//if (k > nb) magmablas_dlaset_batched( MagmaLower, k, k, MAGMA_D_ZERO, MAGMA_D_ZERO, dTstep_array, ldtstep, batchCount, queue );
//TRMV
//T(1:i-1,i) := T(1:i-1,1:i-1) * W(1:i-1) i=[1:k]
// TRMV is split over block of column of size nb
// the update should be done from top to bottom so:
// 1- a gemm using the previous computed columns
// of T to update rectangular upper protion above
// the triangle of my columns
// 2- the columns need to be updated by a serial
// loop over of gemv over itself. since we limit the
// shared memory to nb, this nb column
// are split vertically by chunk of nb rows
dim3 grid(1, 1, batchCount);
for (j=0; j < k; j += nb)
{
prev_n = j;
mycol = min(nb, k-j);
// note that myrow = prev_n + mycol;
if (prev_n > 0 && mycol > 0) {
if (DEBUG == 3) {
printf("doing gemm on the rectangular portion of size %d %d of T(%d,%d)\n",
(int) prev_n, (int) mycol, 0, (int) j );
}
magma_ddisplace_pointers(dW1_displ, dTstep_array, ldtstep, 0, j, batchCount, queue);
magma_ddisplace_pointers(dW2_displ, T_array, ldt, 0, j, batchCount, queue);
magma_dgemm_batched( MagmaNoTrans, MagmaNoTrans,
prev_n, mycol, prev_n,
c_one, T_array, ldt,
dW1_displ, ldtstep,
c_zero, dW2_displ, ldt,
batchCount, queue );
// update my rectangular portion (prev_n,mycol) using sequence of gemv
magma_ddisplace_pointers(dW1_displ, dTstep_array, ldtstep, j, j, batchCount, queue);
magma_ddisplace_pointers(dW3_displ, tau_array, 1, j, 0, batchCount, queue);
for (i=0; i < prev_n; i += nb)
{
rows = min(nb,prev_n-i);
if (DEBUG == 3) {
printf(" doing recdtrmv on the rectangular portion of size %d %d of T(%d,%d)\n",
(int) rows, (int) mycol, (int) i, (int) j );
}
if (rows > 0 && mycol > 0)
{
magma_ddisplace_pointers(dW2_displ, T_array, ldt, i, j, batchCount, queue);
magmablas_dlarft_recdtrmv_sm32x32_batched(rows, mycol, dW3_displ, dW2_displ, ldt, dW1_displ, ldtstep, batchCount, queue);
}
}
}
// the upper rectangular protion is updated, now if needed update the triangular portion
if (stair_T == 0) {
if (DEBUG == 3) {
printf("doing dtrmv on the triangular portion of size %d %d of T(%d,%d)\n",
(int) mycol, (int) mycol, (int) j, (int) j );
}
if (mycol > 0)
{
magma_ddisplace_pointers(dW1_displ, dTstep_array, ldtstep, j, j, batchCount, queue);
magma_ddisplace_pointers(dW3_displ, tau_array, 1, j, 0, batchCount, queue);
magma_ddisplace_pointers(dW2_displ, T_array, ldt, j, j, batchCount, queue);
magmablas_dlarft_dtrmv_sm32x32_batched(mycol, mycol, dW3_displ, dW1_displ, ldtstep, dW2_displ, ldt, batchCount, queue);
}
}
}// end of j
magma_free(dW1_displ);
magma_free(dW2_displ);
magma_free(dW3_displ);
magma_free(dTstep_array);
return 0;
}
extern "C" magma_int_t
magma_dgeqrf_panel_batched(
magma_int_t m, magma_int_t n, magma_int_t nb,
double** dA_array, magma_int_t ldda,
double** tau_array,
double** dT_array, magma_int_t ldt,
double** dR_array, magma_int_t ldr,
double** dW0_displ,
double** dW1_displ,
double *dwork,
double** dW2_displ,
double** dW3_displ,
magma_int_t *info_array,
magma_int_t batchCount, magma_queue_t queue)
{
magma_int_t j, jb;
magma_int_t ldw = nb;
magma_int_t minmn = min(m,n);
for( j=0; j < minmn; j += nb)
{
jb = min(nb, minmn-j);
magma_ddisplace_pointers(dW0_displ, dA_array, ldda, j, j, batchCount, queue);
magma_ddisplace_pointers(dW2_displ, tau_array, 1, j, 0, batchCount, queue);
magma_ddisplace_pointers(dW3_displ, dR_array, ldr, j, j, batchCount, queue); //
//sub-panel factorization
magma_dgeqr2_batched(
m-j, jb,
dW0_displ, ldda,
dW2_displ,
info_array,
batchCount,
queue);
//copy th whole rectangular n,jb from of dA to dR (it's lower portion (which is V's) will be set to zero if needed at the end)
magma_ddisplace_pointers(dW0_displ, dA_array, ldda, 0, j, batchCount, queue);
magma_ddisplace_pointers(dW3_displ, dR_array, ldr, 0, j, batchCount, queue);
magmablas_dlacpy_batched( MagmaFull, minmn, jb, dW0_displ, ldda, dW3_displ, ldr, batchCount, queue );
//set the upper jbxjb portion of V dA(j,j) to 1/0s (note that the rectangular on the top of this triangular of V still non zero but has been copied to dR).
magma_ddisplace_pointers(dW0_displ, dA_array, ldda, j, j, batchCount, queue);
magmablas_dlaset_batched( MagmaUpper, jb, jb, MAGMA_D_ZERO, MAGMA_D_ONE, dW0_displ, ldda, batchCount, queue );
if ( (n-j-jb) > 0) //update the trailing matrix inside the panel
{
magma_dlarft_sm32x32_batched(m-j, jb,
dW0_displ, ldda,
dW2_displ,
dT_array, ldt,
batchCount, queue);
magma_ddisplace_pointers( dW1_displ, dA_array, ldda, j, j + jb, batchCount, queue );
magma_dset_pointer( dW2_displ, dwork, 1, 0, 0, ldw*n, batchCount, queue );
magma_dset_pointer( dW3_displ, dwork + ldw*n*batchCount, 1, 0, 0, ldw*n, batchCount, queue );
magma_dlarfb_gemm_batched(
MagmaLeft, MagmaConjTrans, MagmaForward, MagmaColumnwise,
m-j, n-j-jb, jb,
(const double**)dW0_displ, ldda,
(const double**)dT_array, ldt,
dW1_displ, ldda,
dW2_displ, ldw,
dW3_displ, ldw,
batchCount, queue );
}
}
// copy the remaining portion of dR from dA in case m < n
if ( m < n )
{
magma_ddisplace_pointers(dW0_displ, dA_array, ldda, 0, minmn, batchCount, queue);
magma_ddisplace_pointers(dW3_displ, dR_array, ldr, 0, minmn, batchCount, queue);
magmablas_dlacpy_batched( MagmaFull, minmn, n-minmn, dW0_displ, ldda, dW3_displ, ldr, batchCount, queue );
}
// to be consistent set the whole upper nbxnb of V to 0/1s, in this case no need to set it inside dgeqrf_batched
magma_ddisplace_pointers(dW0_displ, dA_array, ldda, 0, 0, batchCount, queue);
magmablas_dlaset_batched( MagmaUpper, minmn, n, MAGMA_D_ZERO, MAGMA_D_ONE, dW0_displ, ldda, batchCount, queue );
return MAGMA_SUCCESS;
}
