PLASMA  2.4.5
PLASMA - Parallel Linear Algebra for Scalable Multi-core Architectures
 All Data Structures Namespaces Files Functions Variables Typedefs Enumerations Enumerator Macros Groups
Macros | Functions
pdgemm.c File Reference
#include "common.h"
Include dependency graph for pdgemm.c:

Go to the source code of this file.

Macros

#define A(m, n)   BLKADDR(A, double, m, n)
#define B(m, n)   BLKADDR(B, double, m, n)
#define C(m, n)   BLKADDR(C, double, m, n)

Functions

void plasma_pdgemm (plasma_context_t *plasma)
void plasma_pdgemm_quark (PLASMA_enum transA, PLASMA_enum transB, double alpha, PLASMA_desc A, PLASMA_desc B, double beta, PLASMA_desc C, PLASMA_sequence *sequence, PLASMA_request *request)

Detailed Description

PLASMA auxiliary routines PLASMA is a software package provided by Univ. of Tennessee, Univ. of California Berkeley and Univ. of Colorado Denver

Version:
2.4.5
Author:
Emmanuel Agullo
Mathieu Faverge
Date:
2010-11-15 d Tue Nov 22 14:35:39 2011

Definition in file pdgemm.c.

Macro Definition Documentation

#define A (   m,
  n 
)    BLKADDR(A, double, m, n)

Definition at line 18 of file pdgemm.c.

#define B (   m,
  n 
)    BLKADDR(B, double, m, n)

Definition at line 19 of file pdgemm.c.

#define C (   m,
  n 
)    BLKADDR(C, double, m, n)

Definition at line 20 of file pdgemm.c.

Function Documentation

void plasma_pdgemm ( plasma_context_t *  plasma)

Parallel tile matrix-matrix multiplication - static scheduling

Definition at line 24 of file pdgemm.c.

References A, B, BLKLDD, C, CORE_dgemm(), plasma_desc_t::m, plasma_desc_t::mb, plasma_desc_t::mt, plasma_desc_t::n, plasma_desc_t::nb, plasma_desc_t::nt, PLASMA_RANK, PLASMA_SIZE, PLASMA_SUCCESS, plasma_unpack_args_9, PlasmaNoTrans, and plasma_sequence_t::status.

{
PLASMA_enum transA;
PLASMA_enum transB;
double alpha;
PLASMA_desc A;
PLASMA_desc B;
double beta;
PLASMA_desc C;
PLASMA_sequence *sequence;
PLASMA_request *request;
int K, X, Y;
int k, m, n;
int next_m;
int next_n;
int ldam, ldak, ldbn, ldbk, ldcm;
double zbeta;
double zone = (double)1.0;
plasma_unpack_args_9(transA, transB, alpha, A, B, beta, C, sequence, request);
if (sequence->status != PLASMA_SUCCESS)
return;
n = 0;
m = PLASMA_RANK;
while (m >= C.mt && n < C.nt) {
n++;
m = m-C.mt;
}
while (n < C.nt) {
next_m = m;
next_n = n;
next_m += PLASMA_SIZE;
while (next_m >= C.mt && next_n < C.nt) {
next_n++;
next_m = next_m - C.mt;
}
X = m == C.mt-1 ? C.m - m*C.mb : C.mb;
Y = n == C.nt-1 ? C.n - n*C.nb : C.nb;
ldcm = BLKLDD(C, m);
/*
* A: PlasmaNoTrans / B: PlasmaNoTrans
*/
if (transA == PlasmaNoTrans) {
ldam = BLKLDD(A, m);
if (transB == PlasmaNoTrans) {
for (k = 0; k < A.nt; k++) {
K = k == A.nt-1 ? A.n-k*A.nb : A.nb;
ldbk = BLKLDD(B, k);
zbeta = k == 0 ? beta : zone;
CORE_dgemm(
transA, transB,
X, Y, K,
alpha, A(m, k), ldam,
B(k, n), ldbk,
zbeta, C(m, n), ldcm);
}
}
/*
* A: PlasmaNoTrans / B: Plasma[Conj]Trans
*/
else {
ldbn = BLKLDD(B, n);
for (k = 0; k < A.nt; k++) {
K = k == A.nt-1 ? A.n-k*A.nb : A.nb;
zbeta = k == 0 ? beta : zone;
CORE_dgemm(
transA, transB,
X, Y, K,
alpha, A(m, k), ldam,
B(n, k), ldbn,
zbeta, C(m, n), ldcm);
}
}
}
/*
* A: Plasma[Conj]Trans / B: PlasmaNoTrans
*/
else {
if (transB == PlasmaNoTrans) {
for (k = 0; k < A.mt; k++) {
K = k == A.mt-1 ? A.m-k*A.mb : A.mb;
ldak = BLKLDD(A, k);
ldbk = BLKLDD(B, k);
zbeta = k == 0 ? beta : zone;
CORE_dgemm(
transA, transB,
X, Y, K,
alpha, A(k, m), ldak,
B(k, n), ldbk,
zbeta, C(m, n), ldcm);
}
}
/*
* A: Plasma[Conj]Trans / B: Plasma[Conj]Trans
*/
else {
ldbn = BLKLDD(B, n);
for (k = 0; k < A.mt; k++) {
K = k == A.mt-1 ? A.m-k*A.mb : A.mb;
ldak = BLKLDD(A, k);
zbeta = k == 0 ? beta : zone;
CORE_dgemm(
transA, transB,
X, Y, K,
alpha, A(k, m), ldak,
B(n, k), ldbn,
zbeta, C(m, n), ldcm);
}
}
}
m = next_m;
n = next_n;
}
}

Here is the call graph for this function:

Here is the caller graph for this function:

void plasma_pdgemm_quark ( PLASMA_enum  transA,
PLASMA_enum  transB,
double  alpha,
PLASMA_desc  A,
PLASMA_desc  B,
double  beta,
PLASMA_desc  C,
PLASMA_sequence *  sequence,
PLASMA_request *  request 
)

Parallel tile matrix-matrix multiplication - dynamic scheduling

Definition at line 149 of file pdgemm.c.

References B, BLKLDD, C, plasma_desc_t::m, plasma_desc_t::mb, plasma_desc_t::mt, plasma_desc_t::n, plasma_desc_t::nb, plasma_desc_t::nt, plasma_context_self(), PLASMA_SUCCESS, PlasmaNoTrans, plasma_context_struct::quark, QUARK_CORE_dgemm(), plasma_sequence_t::quark_sequence, QUARK_Task_Flag_Set(), Quark_Task_Flags_Initializer, plasma_sequence_t::status, and TASK_SEQUENCE.

{
plasma_context_t *plasma;
Quark_Task_Flags task_flags = Quark_Task_Flags_Initializer;
int m, n, k;
int ldam, ldak, ldbn, ldbk, ldcm;
int tempmm, tempnn, tempkn, tempkm;
double zbeta;
double zone = (double)1.0;
plasma = plasma_context_self();
if (sequence->status != PLASMA_SUCCESS)
return;
QUARK_Task_Flag_Set(&task_flags, TASK_SEQUENCE, (intptr_t)sequence->quark_sequence);
for (m = 0; m < C.mt; m++) {
tempmm = m == C.mt-1 ? C.m-m*C.mb : C.mb;
ldcm = BLKLDD(C, m);
for (n = 0; n < C.nt; n++) {
tempnn = n == C.nt-1 ? C.n-n*C.nb : C.nb;
/*
* A: PlasmaNoTrans / B: PlasmaNoTrans
*/
if (transA == PlasmaNoTrans) {
ldam = BLKLDD(A, m);
if (transB == PlasmaNoTrans) {
for (k = 0; k < A.nt; k++) {
tempkn = k == A.nt-1 ? A.n-k*A.nb : A.nb;
ldbk = BLKLDD(B, k);
zbeta = k == 0 ? beta : zone;
QUARK_CORE_dgemm(
plasma->quark, &task_flags,
transA, transB,
tempmm, tempnn, tempkn, A.mb,
alpha, A(m, k), ldam, /* lda * Z */
B(k, n), ldbk, /* ldb * Y */
zbeta, C(m, n), ldcm); /* ldc * Y */
}
}
/*
* A: PlasmaNoTrans / B: Plasma[Conj]Trans
*/
else {
ldbn = BLKLDD(B, n);
for (k = 0; k < A.nt; k++) {
tempkn = k == A.nt-1 ? A.n-k*A.nb : A.nb;
zbeta = k == 0 ? beta : zone;
QUARK_CORE_dgemm(
plasma->quark, &task_flags,
transA, transB,
tempmm, tempnn, tempkn, A.mb,
alpha, A(m, k), ldam, /* lda * Z */
B(n, k), ldbn, /* ldb * Z */
zbeta, C(m, n), ldcm); /* ldc * Y */
}
}
}
/*
* A: Plasma[Conj]Trans / B: PlasmaNoTrans
*/
else {
if (transB == PlasmaNoTrans) {
for (k = 0; k < A.mt; k++) {
tempkm = k == A.mt-1 ? A.m-k*A.mb : A.mb;
ldak = BLKLDD(A, k);
ldbk = BLKLDD(B, k);
zbeta = k == 0 ? beta : zone;
QUARK_CORE_dgemm(
plasma->quark, &task_flags,
transA, transB,
tempmm, tempnn, tempkm, A.mb,
alpha, A(k, m), ldak, /* lda * X */
B(k, n), ldbk, /* ldb * Y */
zbeta, C(m, n), ldcm); /* ldc * Y */
}
}
/*
* A: Plasma[Conj]Trans / B: Plasma[Conj]Trans
*/
else {
ldbn = BLKLDD(B, n);
for (k = 0; k < A.mt; k++) {
tempkm = k == A.mt-1 ? A.m-k*A.mb : A.mb;
ldak = BLKLDD(A, k);
zbeta = k == 0 ? beta : zone;
QUARK_CORE_dgemm(
plasma->quark, &task_flags,
transA, transB,
tempmm, tempnn, tempkm, A.mb,
alpha, A(k, m), ldak, /* lda * X */
B(n, k), ldbn, /* ldb * Z */
zbeta, C(m, n), ldcm); /* ldc * Y */
}
}
}
}
}
}

Here is the call graph for this function:

Here is the caller graph for this function: