Merge branch 'feature/newalgo/gram' into 'master'

Add a new algorithm for computing a gram matrix

See merge request !152

cmake_modules/local_subs.py

@@ -14,6 +14,7 @@ _extra_blas = [
     ('',                     'she2ge',               'dhe2ge',               'che2ge',               'zhe2ge'              ),
     ('',                     'slatro',               'dlatro',               'clatro',               'zlatro'              ), #=> Replace by getmo/gecmo as in essl
     ('',                     'sbuild',               'dbuild',               'cbuild',               'zbuild'              ), #=> Replace by map function
+    ('',                     'sgram',                'dgram',                'cgram',                'zgram'               ),
 ]
 
 _extra_BLAS = [ [ x.upper() for x in row ] for row in _extra_blas ]

compute/CMakeLists.txt

@@ -242,6 +242,11 @@ set(ZSRC
     ##################
     zbuild.c
     pzbuild.c
+    ##################
+    # Gram
+    ##################
+    zgram.c
+    pzgram.c
 )
 
 precisions_rules_py(CHAMELEON_SRCS_GENERATED "${ZSRC}"

compute/pzgram.c

+/**
+ *
+ * @file pzgram.c
+ *
+ * @copyright 2012-2019 Bordeaux INP, CNRS (LaBRI UMR 5800), Inria,
+ *                      Univ. Bordeaux. All rights reserved.
+ *
+ ***
+ *
+ * @brief Chameleon zgram parallel algorithm
+ *
+ * @version 0.9.2
+ * @author Mathieu Faverge
+ * @author Florent Pruvost
+ * @date 2019-04-10
+ * @precisions normal z -> s d c z
+ *
+ */
+#include "control/common.h"
+
+#define A( m, n )        A,     (m), (n)
+#define W( desc, m, n ) (desc), (m), (n)
+
+static inline void
+chameleon_pzgram_internal( cham_uplo_t uplo,
+                           CHAM_desc_t *A,
+                           CHAM_desc_t *Wcol,
+                           CHAM_desc_t *Welt,
+                           RUNTIME_option_t *options )
+{
+    int m, n;
+    int MT = A->mt;
+    int NT = A->nt;
+    int M  = A->m;
+    int N  = A->n;
+    int P  = Welt->p;
+    int Q  = Welt->q;
+
+    /**
+     *  1) compute (scl,ssq) over columns in each tile
+     */
+    for(n = 0; n < NT; n++) {
+        int mmin = ( uplo == ChamLower ) ? n                      : 0;
+        int mmax = ( uplo == ChamUpper ) ? chameleon_min(n+1, MT) : MT;
+        int tempnn = ( n == (NT-1) ) ? N - n * A->nb : A->nb;
+
+        for(m = mmin; m < mmax; m++) {
+            int tempmm = ( m == (MT-1) ) ? M - m * A->mb : A->mb;
+            int ldam = BLKLDD( A, m );
+
+            if ( n == m ) {
+                INSERT_TASK_dsyssq(
+                    options, ChamColumnwise, uplo, tempmm,
+                    A(m, n), ldam, W( Wcol, m, n) );
+            }
+            else {
+                INSERT_TASK_dgessq(
+                    options, ChamColumnwise, tempmm, tempnn,
+                    A(m, n), ldam, W( Wcol, m, n) );
+                if ( uplo != ChamUpperLower ) {
+                    INSERT_TASK_dgessq(
+                        options, ChamRowwise, tempmm, tempnn,
+                        A(m, n), ldam, W( Wcol, n, m) );
+                }
+            }
+        }
+    }
+
+    for(n = 0; n < NT; n++) {
+        int tempnn = ( n == (NT-1) ) ? N - n * A->nb : A->nb;
+
+        /**
+         *  2) reduce columns (scl,ssq) tiles per processus (between lines)
+         */
+        for(m = P; m < MT; m++) {
+            INSERT_TASK_dplssq(
+                options, ChamColumnwise, 1, tempnn,
+                W( Wcol, m,   n ),
+                W( Wcol, m%P, n ) );
+        }
+
+        /**
+         *  3) reduce columns (scl,ssq) tiles on the first line of tiles
+         */
+        for(m = 1; m < P; m++) {
+            INSERT_TASK_dplssq(
+                options, ChamColumnwise, 1, tempnn,
+                W( Wcol, m, n ),
+                W( Wcol, 0, n ) );
+        }
+
+        /* 4) reduce (scl,ssq) inside each tile of the first line of tiles for the global sum square */
+        INSERT_TASK_dplssq(
+            options, ChamEltwise, 1, tempnn,
+            W( Wcol, 0, n ),
+            W( Welt, 0, n ) );
+
+        /* 5) deduce the sum square for each column from the pairs (scl,ssq) -> sqrt(sum) = scl*sqrt(ssq) */
+        INSERT_TASK_dplssq2( options, tempnn, W( Wcol, 0, n ) );
+    }
+
+    /* 6) reduce global sum squares on each processus (between columns) */
+    for(n = Q; n < NT; n++) {
+        INSERT_TASK_dplssq( options, ChamEltwise, 1, 1, W( Welt, 0, n), W( Welt, 0, n%Q) );
+    }
+
+    /* 7) reduce global sum squares on the first tile (index 0, 0) */
+    for(n = 1; n < Q; n++) {
+        INSERT_TASK_dplssq(
+            options, ChamEltwise, 1, 1, W( Welt, 0, n), W( Welt, 0, 0) );
+    }
+
+    /* 8) deduce the global sum square from the pair (scl,ssq) -> sqrt(sum) = scl*sqrt(ssq) */
+    INSERT_TASK_dplssq2( options, 1, W( Welt, 0, 0) );
+
+    /* Finally compute Gram matrix coefficients inplace */
+    for(n = 0; n < NT; n++) {
+        int mmin = ( uplo == ChamLower ) ? n                      : 0;
+        int mmax = ( uplo == ChamUpper ) ? chameleon_min(n+1, MT) : MT;
+        int tempnn = ( n == (NT-1) ) ? N - n * A->nb : A->nb;
+
+        for(m = mmin; m < mmax; m++) {
+            int tempmm = ( m == (MT-1) ) ? M - m * A->mb : A->mb;
+            int ldam = BLKLDD( A, m );
+
+            INSERT_TASK_zgram(
+                options,
+                ( m == n ) ? uplo : ChamUpperLower,
+                A->m, A->n, tempmm, tempnn,
+                W( Wcol, 0, m ), 2,
+                W( Wcol, 0, n ), 2,
+                W( Welt, 0, 0 ),
+                A( m, n ), ldam );
+        }
+    }
+}
+
+/**
+ *
+ */
+void chameleon_pzgram( cham_uplo_t uplo, CHAM_desc_t *A, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request )
+{
+    CHAM_context_t *chamctxt;
+    RUNTIME_option_t options;
+    CHAM_desc_t Wcol;
+    CHAM_desc_t Welt;
+
+    int workmt, worknt;
+    int m, n, tempmm, tempnn, ldw;
+
+    chamctxt = chameleon_context_self();
+    if ( sequence->status != CHAMELEON_SUCCESS ) {
+        return;
+    }
+    RUNTIME_options_init(&options, chamctxt, sequence, request);
+
+    workmt = chameleon_max( A->mt, A->p );
+    worknt = chameleon_max( A->nt, A->q );
+
+    RUNTIME_options_ws_alloc( &options, 1, 0 );
+
+    chameleon_desc_init( &Wcol, CHAMELEON_MAT_ALLOC_TILE, ChamRealDouble, 2, A->nb, 2*A->nb,
+                         2*workmt, A->n, 0, 0, 2*workmt, A->n, A->p, A->q,
+                         NULL, NULL, NULL );
+
+    chameleon_desc_init( &Welt, CHAMELEON_MAT_ALLOC_TILE, ChamRealDouble, 2, 1, 2,
+                         2, worknt, 0, 0, 2, worknt, A->p, A->q,
+                         NULL, NULL, NULL );
+
+    /* Initialize Wcol */
+    for(m = 0; m < Wcol.mt; m++) {
+        tempmm = m == Wcol.mt-1 ? Wcol.m-m*Wcol.mb : Wcol.mb;
+        ldw = Wcol.get_blkldd(&Wcol, m);
+        for(n = 0; n < Wcol.nt; n++) {
+            tempnn = n == Wcol.nt-1 ? Wcol.n-n*Wcol.nb : Wcol.nb;
+            INSERT_TASK_dlaset(
+                &options,
+                ChamUpperLower, tempmm, tempnn,
+                -1., -1.,
+                W( &Wcol, m, n ), ldw );
+        }
+    }
+    /* Initialize Welt */
+    for(m = 0; m < Welt.mt; m++) {
+        tempmm = m == Welt.mt-1 ? Welt.m-m*Welt.mb : Welt.mb;
+        ldw = Welt.get_blkldd(&Welt, m);
+        for(n = 0; n < Welt.nt; n++) {
+            tempnn = n == Welt.nt-1 ? Welt.n-n*Welt.nb : Welt.nb;
+            INSERT_TASK_dlaset(
+                &options,
+                ChamUpperLower, tempmm, tempnn,
+                -1., -1.,
+                W( &Welt, m, n ), ldw );
+        }
+    }
+
+    chameleon_pzgram_internal( uplo, A, &Wcol, &Welt, &options );
+
+    CHAMELEON_Desc_Flush( &Wcol, sequence );
+    CHAMELEON_Desc_Flush( &Welt, sequence );
+    CHAMELEON_Desc_Flush( A, sequence );
+    RUNTIME_sequence_wait( chamctxt, sequence );
+
+    chameleon_desc_destroy( &Wcol );
+    chameleon_desc_destroy( &Welt );
+
+    RUNTIME_options_ws_free(&options);
+    RUNTIME_options_finalize(&options, chamctxt);
+}

compute/zgram.c

+/**
+ *
+ * @file zgram.c
+ *
+ * @copyright 2012-2019 Bordeaux INP, CNRS (LaBRI UMR 5800), Inria,
+ *                      Univ. Bordeaux. All rights reserved.
+ *
+ ***
+ *
+ * @brief Chameleon zgram wrappers
+ *
+ * @version 0.9.2
+ * @author Mathieu Faverge
+ * @author Florent Pruvost
+ * @date 2019-04-10
+ * @precisions normal z -> s d c z
+ *
+ */
+#include "control/common.h"
+
+/**
+ ********************************************************************************
+ *
+ * @ingroup CHAMELEON_Complex64_t
+ *
+ *  CHAMELEON_zgram replace a general matrix by the Gram matrix inplace
+ *
+ *    \f[
+ *    d_{i.}^2 = (1/n) \sum_j d_{ij}^2` and :math:`d_{..}^2 = (1/n^2) \sum_{i,j} d_{ij}^2 \\
+ *    A_{i,j} = -(1/2) (d_{ij}^2 - d_{i.}^2 - d_{.j}^2 + d_{..}^2)
+ *    \f]
+ *
+ *******************************************************************************
+ *
+ * @param[in] N
+ *          The number of lines and columns of the matrix A. N >= 0. When N = 0,
+ *          the returned value is set to zero.
+ *
+ * @param[in] A
+ *          The N-by-N matrix A.
+ *
+ * @param[in] LDA
+ *          The leading dimension of the array A. LDA >= max(1,N).
+ *
+ *******************************************************************************
+ *
+* @retval CHAMELEON_SUCCESS successful exit
+ *
+ *******************************************************************************
+ *
+ * @sa CHAMELEON_zgram_Tile
+ * @sa CHAMELEON_zgram_Tile_Async
+ * @sa CHAMELEON_sgram
+ *
+ */
+int CHAMELEON_zgram( cham_uplo_t uplo, int N, CHAMELEON_Complex64_t *A, int LDA )
+{
+    int NB;
+    int status;
+    CHAM_context_t *chamctxt;
+    RUNTIME_sequence_t *sequence = NULL;
+    RUNTIME_request_t request = RUNTIME_REQUEST_INITIALIZER;
+    CHAM_desc_t descAl, descAt;
+
+    chamctxt = chameleon_context_self();
+    if (chamctxt == NULL) {
+        chameleon_fatal_error("CHAMELEON_zgram", "CHAMELEON not initialized");
+        return CHAMELEON_ERR_NOT_INITIALIZED;
+    }
+    /* Check input arguments */
+    if (N < 0) {
+        chameleon_error("CHAMELEON_zgram", "illegal value of N");
+        return -1;
+    }
+    if (LDA < chameleon_max(1, N)) {
+        chameleon_error("CHAMELEON_zgram", "illegal value of LDA");
+        return -3;
+    }
+
+    /* Quick return */
+    if (N == 0)
+        return CHAMELEON_SUCCESS;
+
+    /* Tune NB depending on M, N & NRHS; Set NBNB */
+    status = chameleon_tune(CHAMELEON_FUNC_DGEMM, N, N, 0);
+    if (status != CHAMELEON_SUCCESS) {
+        chameleon_error("CHAMELEON_zgram", "chameleon_tune() failed");
+        return status;
+    }
+
+    /* Set NT */
+    NB = CHAMELEON_NB;
+
+    chameleon_sequence_create( chamctxt, &sequence );
+
+    /* Submit the matrix conversion */
+    chameleon_zlap2tile( chamctxt, &descAl, &descAt, ChamDescInout, uplo,
+                         A, NB, NB, LDA, N, N, N, sequence, &request );
+
+    /* Call the tile interface */
+    CHAMELEON_zgram_Tile_Async( uplo, &descAt, sequence, &request );
+
+    /* Submit the matrix conversion back */
+    chameleon_ztile2lap( chamctxt, &descAl, &descAt,
+                         ChamDescInout, uplo, sequence, &request );
+
+    chameleon_sequence_wait( chamctxt, sequence );
+
+    /* Cleanup the temporary data */
+    chameleon_ztile2lap_cleanup( chamctxt, &descAl, &descAt );
+
+    status = sequence->status;
+    chameleon_sequence_destroy( chamctxt, sequence );
+    return status;
+}
+
+/**
+ ********************************************************************************
+ *
+ * @ingroup CHAMELEON_Complex64_t_Tile
+ *
+ *  CHAMELEON_zgram_Tile - Tile equivalent of CHAMELEON_zgram().
+ *  Operates on matrices stored by tiles.
+ *  All matrices are passed through descriptors.
+ *  All dimensions are taken from the descriptors.
+ *
+ *******************************************************************************
+ *
+ * @param[in] A
+ *          The N-by-N matrix A.
+ *
+ *******************************************************************************
+ *
+ * @retval CHAMELEON_SUCCESS successful exit
+ *
+ *******************************************************************************
+ *
+ * @sa CHAMELEON_zgram
+ * @sa CHAMELEON_zgram_Tile_Async
+ * @sa CHAMELEON_sgram_Tile
+ *
+ */
+int CHAMELEON_zgram_Tile( cham_uplo_t uplo, CHAM_desc_t *A )
+{
+    CHAM_context_t *chamctxt;
+    RUNTIME_sequence_t *sequence = NULL;
+    RUNTIME_request_t request = RUNTIME_REQUEST_INITIALIZER;
+    int status;
+
+    chamctxt = chameleon_context_self();
+    if (chamctxt == NULL) {
+        chameleon_fatal_error("CHAMELEON_zgram_Tile", "CHAMELEON not initialized");
+        return CHAMELEON_ERR_NOT_INITIALIZED;
+    }
+    chameleon_sequence_create( chamctxt, &sequence );
+
+    CHAMELEON_zgram_Tile_Async( uplo, A, sequence, &request );
+
+    CHAMELEON_Desc_Flush( A, sequence );
+
+    chameleon_sequence_wait( chamctxt, sequence );
+    status = sequence->status;
+    chameleon_sequence_destroy( chamctxt, sequence );
+    return status;
+}
+
+/**
+ ********************************************************************************
+ *
+ * @ingroup CHAMELEON_Complex64_t_Tile_Async
+ *
+ *  CHAMELEON_zgram_Tile_Async - Non-blocking equivalent of CHAMELEON_zgram_Tile().
+ *  May return before the computation is finished.
+ *  Allows for pipelining of operations at runtime.
+ *
+ *******************************************************************************
+ *
+ * @param[in] sequence
+ *          Identifies the sequence of function calls that this call belongs to
+ *          (for completion checks and exception handling purposes).
+ *
+ * @param[out] request
+ *          Identifies this function call (for exception handling purposes).
+ *
+ *******************************************************************************
+ *
+ * @sa CHAMELEON_zgram
+ * @sa CHAMELEON_zgram_Tile
+ * @sa CHAMELEON_sgram_Tile_Async
+ *
+ */
+int CHAMELEON_zgram_Tile_Async( cham_uplo_t uplo, CHAM_desc_t *A,
+                                RUNTIME_sequence_t *sequence, RUNTIME_request_t *request )
+{
+    CHAM_context_t *chamctxt;
+
+    chamctxt = chameleon_context_self();
+    if (chamctxt == NULL) {
+        chameleon_fatal_error("CHAMELEON_zgram_Tile", "CHAMELEON not initialized");
+        return CHAMELEON_ERR_NOT_INITIALIZED;
+    }
+    if (sequence == NULL) {
+        chameleon_fatal_error("CHAMELEON_zgram_Tile", "NULL sequence");
+        return CHAMELEON_ERR_UNALLOCATED;
+    }
+    if (request == NULL) {
+        chameleon_fatal_error("CHAMELEON_zgram_Tile", "NULL request");
+        return CHAMELEON_ERR_UNALLOCATED;
+    }
+    /* Check sequence status */
+    if (sequence->status == CHAMELEON_SUCCESS) {
+        request->status = CHAMELEON_SUCCESS;
+    }
+    else {
+        return chameleon_request_fail(sequence, request, CHAMELEON_ERR_SEQUENCE_FLUSHED);
+    }
+
+    /* Check descriptors for correctness */
+    if (chameleon_desc_check(A) != CHAMELEON_SUCCESS) {
+        chameleon_error("CHAMELEON_zgram_Tile", "invalid descriptor");
+        return chameleon_request_fail(sequence, request, CHAMELEON_ERR_ILLEGAL_VALUE);
+    }
+    /* Check input arguments */
+    if (A->nb != A->mb) {
+        chameleon_error("CHAMELEON_zgram_Tile", "only square tiles supported");
+        return chameleon_request_fail(sequence, request, CHAMELEON_ERR_ILLEGAL_VALUE);
+    }
+    /* Quick return */
+    if (chameleon_min(A->m, A->n) == 0) {
+        return CHAMELEON_SUCCESS;
+    }
+
+    chameleon_pzgram( uplo, A, sequence, request );
+
+    return CHAMELEON_SUCCESS;
+}

control/compute_z.h

@@ -115,7 +115,10 @@ void chameleon_pzungqr_param( int genD, int K, const libhqr_tree_t *qrtree,
                               CHAM_desc_t *TS, CHAM_desc_t *TT, CHAM_desc_t *D,
                               RUNTIME_sequence_t *sequence, RUNTIME_request_t *request);
 
-
+/**
+ * Gram function prototypes
+ */
+void chameleon_pzgram( cham_uplo_t uplo, CHAM_desc_t *A, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request );
 
 /**
  *  LAPACK/Tile Descriptor accesses

coreblas/compute/CMakeLists.txt

@@ -44,6 +44,7 @@ set(ZSRC
     core_zgetrf.c
     core_zgetrf_incpiv.c
     core_zgetrf_nopiv.c
+    core_zgram.c
     core_zhe2ge.c
     core_zherfb.c
     core_zhemm.c

coreblas/compute/core_zgessq.c

@@ -174,6 +174,24 @@ int CORE_zgessq( cham_store_t storev, int M, int N,
                  const CHAMELEON_Complex64_t *A, int LDA,
                  double *sclssq )
 {
+    int i;
+    int K;
+
+    /* initialize pairs scale, sumsquare if not already done */
+    if ( storev == ChamColumnwise ) {
+        K = N;
+    } else if ( storev == ChamRowwise ) {
+        K = M;
+    } else {
+        K = 1;
+    }
+    for (i=0; i<2*K; i+=2) {
+        if ( ( sclssq[i] == -1. ) && ( sclssq[i+1] == -1. ) ) {
+            sclssq[i] = 1.;
+            sclssq[i+1] = 0.;
+        }
+    }
+
     if (storev == ChamColumnwise) {
         CORE_zgessq_col( M, N, A, LDA, sclssq );
     }

coreblas/compute/core_zgram.c

+/**
+ *
+ * @file core_zgram.c
+ *
+ * @copyright 2012-2019 Bordeaux INP, CNRS (LaBRI UMR 5800), Inria,
+ *                      Univ. Bordeaux. All rights reserved.
+ *
+ ***
+ *
+ * @brief Chameleon core_zgram CPU kernel
+ *
+ * @version 0.9.2
+ * @author Mathieu Faverge
+ * @author Florent Pruvost
+ * @date 2019-04-10
+ * @precisions normal z -> s d c z
+ *
+ */
+#include "coreblas.h"
+
+/**
+ *******************************************************************************
+ *
+ * @ingroup CORE_CHAMELEON_Complex64_t
+ *
+ *  CORE_zgram computes the Gram matrix factors of A inplace
+ *
+ *    \f[
+ *    d_{i.}^2 = (1/n) \sum_j d_{ij}^2` and :math:`d_{..}^2 = (1/n^2) \sum_{i,j} d_{ij}^2 \\
+ *    A_{i,j} = -(1/2) (d_{ij}^2 - d_{i.}^2 - d_{.j}^2 + d_{..}^2)
+ *    \f]
+ *
+ *******************************************************************************
+ *
+ * @param[in] uplo
+ *          Specifies the part of the matrix A to be computed.
+ *            = ChamUpperLower: All the matrix A
+ *            = ChamUpper: Upper triangular part
+ *            = ChamLower: Lower triangular part
+ *
+ * @param[in] M
+ *          The number of rows of the overall matrix.  M >= 0.
+ *
+ * @param[in] N
+ *         The number of columns of the overall matrix.  N >= 0.
+ *
+ * @param[in] Mt
+ *          The number of rows of the tile A.  Mt >= 0.
+ *
+ * @param[in] Nt
+ *         The number of columns of the tile A.  Nt >= 0.
+ *
+ * @param[in] Di
+ *         The 1-by-Mt tile containing the sum of squares by rows.
+ *
+ * @param[in] LDDI
+ *         The leading dimension of the array Di.
+ *
+ * @param[in] Dj
+ *         The 1-by-Nt tile containing the sum of squares by columns.
+ *
+ * @param[in] LDDJ
+ *         The leading dimension of the array Dj.
+ *
+ * @param[in] D
+ *         The sum of squares of all the matrix.
+ *
+ * @param[in,out] A
+ *         On entry, the M-by-N tile A.
+ *         On exit, updated by the application of L.
+ *
+ * @param[in] LDA
+ *         The leading dimension of the array A.  LDA >= max(1,M).
+ *
+ *******************************************************************************
+ *
+ * @retval CHAMELEON_SUCCESS successful exit
+ * @retval <0 if INFO = -k, the k-th argument had an illegal value
+ *
+ */
+
+int CORE_zgram( cham_uplo_t uplo,
+                int M, int N,
+                int Mt, int Nt,
+                const double *Di, int LDDI,
+                const double *Dj, int LDDJ,
+                const double *D,
+                double *A, int LDA )
+{
+    int i, j;
+    double coef = -0.5;
+    double di, dj, d;
+
+    /* Check input arguments */
+    if (uplo != ChamUpper && uplo != ChamLower && uplo != ChamUpperLower ) {
+        coreblas_error(1, "Illegal value of uplo");
+        return -1;
+    }
+    if (M < 0) {
+        coreblas_error(2, "Illegal value of M");
+        return -2;
+    }
+    if (N < 0) {
+        coreblas_error(3, "Illegal value of N");
+        return -3;
+    }
+    if (Mt < 0) {
+        coreblas_error(4, "Illegal value of Mt");
+        return -4;
+    }
+    if (Nt < 0) {
+        coreblas_error(5, "Illegal value of Nt");
+        return -5;
+    }
+    if ((LDA < chameleon_max(1,Mt)) && (Mt > 0)) {
+        coreblas_error(12, "Illegal value of LDA");
+        return -12;
+    }
+