diff --git a/src/masterworker3.cpp b/src/masterworker3.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..418f715b23dfb2d0d105fca09b309e4b6703eadf
--- /dev/null
+++ b/src/masterworker3.cpp
@@ -0,0 +1,222 @@
+#include <iostream>
+#include <mpi.h>
+#include <ginac/ginac.h>
+
+#include "products.h"
+#include "utils_parall.h"
+#include "parall_constants.h"
+#include "parall_internal.h"
+#include "utils.h"
+
+namespace gi = GiNaC;
+
+/*******************************************************************************
+ * Parallel 1-level decomposition *
+ *******************************************************************************/
+
+gi::ex multiply_1level_master3( tensor3D_t& T, unsigned int size, MPI_Comm comm = MPI_COMM_WORLD ) {
+ gi::ex Tens = 0;
+ unsigned int a2, a4;
+ gi::lst symbols;
+
+ MPI_Status status;
+ char* expr_c;
+ size_t expr_c_size = 0;
+ int src, np, running = 0;
+ unsigned int len;
+ parameters_2_1_t pzero( 0, 0 );
+
+ MPI_Comm_size( comm, &np );
+
+ expr_c = NULL;
+ expr_c = (char*) malloc( 3279 ); // TMP
+
+ int i, j;
+ i = 0;
+ j = 0;
+
+ int receivedresults = 0;
+ unsigned int N = size/2;
+
+ std::vector<parameters_2_1_t> input;
+ std::vector<std::string> results_s;
+ std::vector<gi::ex> results;
+
+ /* Build a list of argument sets */
+
+ for( a4 = 0 ; a4 < N ; a4++ ){
+ for( a2 = 0; a2 < N ; a2++ ){
+ parameters_2_1_t p( a4, a2 );
+ input.push_back( p );
+ }
+ }
+
+ /* Compute the set of symbols */
+ /* Could be done while the first slave is working */
+
+ symbols = all_symbols_3D( size );
+
+ /* Distribute the work */
+
+ while( input.size() > 0 ) {
+ MPI_Recv( &len, 1, MPI_UNSIGNED, MPI_ANY_SOURCE, MPI_ANY_TAG, comm, &status );
+ src = status.MPI_SOURCE;
+
+ if( status.MPI_TAG == TAG_PULL ) {
+
+ /* Nothing else will come: just send wome work */
+ send_work( input, src, comm );
+
+ } else {
+ if( status.MPI_TAG == TAG_RES ){
+ src = status.MPI_SOURCE;
+
+ /* The first message contains the length of what is coming next */
+ if( len != 0 ) {
+ if( len > expr_c_size ) {
+ expr_c_size = len;
+ if( NULL != expr_c ) free( expr_c );
+ expr_c = (char*)malloc( expr_c_size ); // The \0 was added by the slave
+ }
+
+ /* Receive the result */
+ MPI_Recv( expr_c, len, MPI_CHAR, src, TAG_EXPR, comm, &status );
+
+ /* put it in the result queue */
+ std::string s( expr_c );
+
+ send_work( input, src, comm );
+
+ /* Process what I have just received */
+ /* Could be given to a slave... */
+ gi::ex received = de_linearize_expression( s, symbols );
+ Tens += received;
+#if DEBUG
+ results.push_back( received );
+ results_s.push_back( s );
+ receivedresults++;
+#endif
+ } else {
+ /* Send more work */
+ send_work( input, src, comm );
+ }
+ } else{
+ std::cerr << "Wrong tag received " << status.MPI_TAG << std::endl;
+ }
+ }
+ }
+
+ /* Wait until everyone is done */
+
+ running = np - 1; // all the slaves are running
+ while( running > 0 ) {
+
+ MPI_Recv( &len, 1, MPI_UNSIGNED, MPI_ANY_SOURCE, MPI_ANY_TAG, comm, &status );
+ src = status.MPI_SOURCE;
+
+ if( len != 0 ) {
+ if( len > expr_c_size ) {
+ expr_c_size = len;
+ if( NULL != expr_c ) free( expr_c );
+ expr_c = (char*)malloc( expr_c_size ); // The \0 was added by the slave
+ }
+
+ /* Receive the result */
+ MPI_Recv( expr_c, len, MPI_CHAR, src, TAG_EXPR, comm, &status );
+
+ /* And send the END signal */
+ send_end( src, pzero, comm );
+ running--;
+
+ /* Process what I have just received */
+ /* Could be given to a slave... */
+ /* put it in the result queue */
+ std::string s( expr_c );
+ gi::ex received = de_linearize_expression( s, symbols );
+ Tens += received;
+#if DEBUG
+ results.push_back( received );
+ results_s.push_back( s );
+ receivedresults++;
+#endif
+ } else {
+ send_end( src, pzero, comm );
+ running--;
+ }
+ }
+
+#if DEBUG
+ std::cout << "Received " << receivedresults << " results" << std::endl;
+
+ std::cout << "Tpara=" << Tens << ";" << std::endl;
+#endif
+
+ if( NULL != expr_c) free( expr_c );
+ return Tens;
+}
+
+void multiply_1level_slave3( tensor3D_t& T, int size, MPI_Comm comm = MPI_COMM_WORLD ) {
+ gi::ex Tens;
+ unsigned int a2, a4;
+ unsigned int len = 0;
+
+ parameters_t params;
+ MPI_Status status;
+ char* expr_c;
+
+ int rank;
+ MPI_Comm_rank( comm, &rank );
+
+ /* Ask for some work */
+
+ MPI_Send( &len, 1, MPI_UNSIGNED, ROOT, TAG_PULL, comm );
+
+ while( true ){
+ /* Receive a set of parameters */
+
+ MPI_Recv( ¶ms, 1, DT_PARAMETERS_2_1, ROOT, MPI_ANY_TAG, comm, &status );
+
+ if( status.MPI_TAG == TAG_WORK ){
+ a4 = params.a4;
+ a2 = params.a2;
+
+ Tens = one_level1_product( &T, size, a4, a2 );
+ send_result( Tens );
+
+ } else {
+ if( status.MPI_TAG == TAG_END ){
+ return;
+ } else {
+ std::cerr << "Wrong tag received on slave " << status.MPI_TAG << std::endl;
+ }
+ }
+ }
+}
+
+/* Communication protocol:
+ M -> W: always the same size, therefore unique communication
+ W -> M: send an unsigned int (size of the expression), then the expression (table of chars)
+*/
+
+gi::ex multiply_1level_mw3( tensor3D_t& T, int size ) { // simpler: same dimension everywhere
+ int rank;
+ gi::ex Tens = 0;
+ MPI_Comm_rank( MPI_COMM_WORLD, &rank );
+
+ /* Create a new datatype for the parameters */
+
+ create_parameters_datatype_2_1();
+
+ /* Here we go */
+
+ if( 0 == rank ) {
+ Tens = multiply_1level_master3( T, size );
+ } else {
+ multiply_1level_slave3( T, size );
+ }
+
+ /* Finalize */
+
+ free_parameters_2_1_dt();
+ return Tens;
+}
diff --git a/src/mw_addslave3.cpp b/src/mw_addslave3.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..bd1761ede1a0eda01100611335d27dfe95326d98
--- /dev/null
+++ b/src/mw_addslave3.cpp
@@ -0,0 +1,237 @@
+#include <iostream>
+#include <mpi.h>
+#include <ginac/ginac.h>
+
+#include "products.h"
+#include "utils_parall.h"
+#include "parall_constants.h"
+#include "parall_internal.h"
+#include "utils.h"
+#include "profiling.h"
+
+namespace gi = GiNaC;
+
+/*******************************************************************************
+ * Parallel 1-level decomposition with addition on a slave *
+ *******************************************************************************/
+
+gi::ex multiply_1level_master_addslave3( tensor3D_t& T, unsigned int size, MPI_Comm comm = MPI_COMM_WORLD ) {
+ gi::ex Tens = 0;
+ unsigned int a2, a4;
+ gi::ex A;
+ gi::lst symbols;
+
+ MPI_Status status;
+ char* expr_c;
+ size_t expr_c_size = 0;
+ int src, np, running = 0;
+ unsigned int len;
+ parameters_2_1_t pzero( 0, 0 );
+
+ MPI_Comm_size( comm, &np );
+
+ expr_c = NULL;
+ expr_c = (char*) malloc( 3279 );
+
+ int i, j;
+ i = 0;
+ j = 0;
+
+ int receivedresults = 0;
+ unsigned int N = size/2;
+
+ std::vector<parameters_2_1_t> input;
+ std::vector<std::string> results; /* length and char* */
+
+ /* Build a list of argument sets */
+
+ for( a4 = 0 ; a4 < N ; a4++ ){
+ for( a2 = 0; a2 < N ; a2++ ){
+ parameters_2_1_t p( a4, a2 );
+ input.push_back( p );
+ }
+ }
+
+ /* Compute the set of symbols */
+ /* Could be done while the first slave is working */
+
+ symbols = all_symbols_3D( size );
+
+ /* Distribute the work */
+
+ while( input.size() > 0 ) {
+ MPI_Recv( &len, 1, MPI_UNSIGNED, MPI_ANY_SOURCE, MPI_ANY_TAG, comm, &status );
+
+ if( status.MPI_TAG == TAG_PULL ) {
+ /* Nothing else will come: just send wome work */
+ src = status.MPI_SOURCE;
+ send_work( input, src );
+
+ } else {
+ if( status.MPI_TAG == TAG_RES ){
+ src = status.MPI_SOURCE;
+
+ /* The first message contains the length of what is coming next */
+ if( len != 0 ) {
+ if( len > expr_c_size ) {
+ expr_c_size = len;
+ if( NULL != expr_c ) free( expr_c );
+ expr_c = (char*)malloc( expr_c_size ); // The \0 was added by the slave
+ }
+
+ /* Receive the result */
+ MPI_Recv( expr_c, len, MPI_CHAR, src, TAG_EXPR, comm, &status );
+
+ /* Put it in the result queue */
+ results.push_back( std::string( expr_c ) );
+ }
+
+ /* Send more work */
+ send_work_addslave( input, results, src );
+ } else {
+ std::cerr << "Wrong tag received " << status.MPI_TAG << std::endl;
+ }
+
+ }
+ }
+
+ /* Wait until everyone is done */
+
+ running = np - 1; // all the slaves are running
+ while( running > 0 ) {
+ /* TODO: here all we should receive is either TAG_EXPR or TAG_PULL if the input is too small */
+ MPI_Recv( &len, 1, MPI_UNSIGNED, MPI_ANY_SOURCE, MPI_ANY_TAG, comm, &status );
+ src = status.MPI_SOURCE;
+
+ if( len != 0 ) {
+ if( len > expr_c_size ) {
+ expr_c_size = len;
+ if( NULL != expr_c ) free( expr_c );
+ expr_c = (char*)malloc( expr_c_size ); // The \0 was added by the slave
+ }
+
+ /* Receive the result */
+ MPI_Recv( expr_c, len, MPI_CHAR, src, TAG_EXPR, comm, &status );
+
+ /* Put it in the result queue */
+ results.push_back( std::string( expr_c ) );
+ }
+ send_add_or_end_addslave( results, src, &running, pzero );
+ }
+
+ /* Add whatever I have left */
+ Tens = add_expressions( results, symbols );
+
+#if DEBUG
+ std::cout << "Received " << receivedresults << " results" << std::endl;
+
+ std::cout << "Tpara=" << Tens << ";" << std::endl;
+#endif
+
+ if( NULL != expr_c) free( expr_c );
+ return Tens;
+}
+
+void multiply_1level_slave_addslave3( tensor3D_t& T, unsigned int size, MPI_Comm comm = MPI_COMM_WORLD ) {
+ gi::ex Tens;
+ int a2, a4;
+ unsigned int len = 0;
+
+ parameters_t params;
+ MPI_Status status;
+ char* expr_c;
+
+ int rank;
+ MPI_Comm_rank( comm, &rank );
+
+ /* Ask for some work */
+
+ MPI_Send( &len, 1, MPI_UNSIGNED, ROOT, TAG_PULL, comm );
+
+ /* Compute the set of symbols */
+
+ gi::lst symbols = all_symbols_3D( size );
+
+ while( true ){
+ /* Receive a set of parameters */
+
+ MPI_Recv( ¶ms, 1, DT_PARAMETERS_2_1, ROOT, MPI_ANY_TAG, comm, &status );
+
+ if( status.MPI_TAG == TAG_WORK ){
+ a4 = params.a4;
+ a2 = params.a2;
+
+ Tens = one_level1_product( &T, size, a4, a2 );
+ send_result( Tens );
+
+ } else {
+ if( status.MPI_TAG == TAG_ADD ) {
+ /* Receive a set of expressions to add */
+
+ /* Number of expressions received */
+ int nb = params.a4;
+
+ /* Length of each string */
+
+ unsigned int* lengths = (unsigned int*) malloc( nb*sizeof( unsigned int ) );
+ MPI_Recv( lengths, nb, MPI_INT, ROOT, TAG_ADD, comm, &status );
+ std::vector<std::string> results_s;
+ char* c_str;
+ int i;
+ int len;
+ for( i = 0 ; i < nb ; i++ ) {
+ len = lengths[i] + 1;
+ c_str = (char*) malloc( len );
+ MPI_Recv( c_str, len, MPI_CHAR, ROOT, TAG_ADD, comm, &status );
+ c_str[len-1] = '\0'; // The master sends C++ strings, which do not contain the final '\0'
+ results_s.push_back( std::string( c_str ) );
+ free( c_str );
+ }
+
+ /* Delinearize all the expressions and add them */
+
+ Tens = add_expressions( results_s, symbols );
+
+ /* Send the result */
+
+ send_result( Tens );
+
+ } else {
+ if( status.MPI_TAG == TAG_END ){
+ return;
+ } else {
+ std::cerr << "Wrong tag received on slave " << status.MPI_TAG << std::endl;
+ }
+ }
+ }
+ }
+}
+
+/* Communication protocol:
+ M -> W: always the same size, therefore unique communication
+ W -> M: send an unsigned int (size of the expression), then the expression (table of chars)
+*/
+
+gi::ex multiply_1level_mw_addslave3( tensor3D_t& T, int size ) { // simpler: same dimension everywhere
+ int rank;
+ gi::ex Tens = 0;
+ MPI_Comm_rank( MPI_COMM_WORLD, &rank );
+
+ /* Create a new datatype for the parameters */
+
+ create_parameters_datatype_2_1();
+
+ /* Here we go */
+
+ if( 0 == rank ) {
+ Tens = multiply_1level_master_addslave3( T, size );
+ } else {
+ multiply_1level_slave_addslave3( T, size );
+ }
+
+ /* Finalize */
+
+ free_parameters_2_1_dt();
+ return Tens;
+}
+