diff --git a/src/Makefile b/src/Makefile
index a4eefda9d85f2267223b27af0e071049ae633546..cb3cdf2b1c47c34c7a3f8bae8ad35b4d2116399b 100644
--- a/src/Makefile
+++ b/src/Makefile
@@ -28,8 +28,8 @@ NP = 5
MPISRC = masterworker.cpp mw_addslave.cpp \
perf.cpp sequential.cpp tensormatrix_mpi.cpp \
- utils.cpp utils_parall.cpp profiling.cpp
-#mw_combined.cpp hierarchical.cpp
+ utils.cpp utils_parall.cpp profiling.cpp mw_combined.cpp
+# hierarchical.cpp
MPIOBJ= $(MPISRC:.cpp=.o)
diff --git a/src/mw_combined.cpp b/src/mw_combined.cpp
index b6bc69eb49180787bf239b7d71fe4fffa3efd0d4..663eda0cd765851901bb00b5c0e5359bc82ba4ce 100644
--- a/src/mw_combined.cpp
+++ b/src/mw_combined.cpp
@@ -25,8 +25,8 @@ typedef enum {
/* This one is a "regular" master. It returns either when it is done, or when it decides to switch to another algorithm.
*/
-end_code_t multiply_combined_master_initial( tensor3D_t& T, matrix_int_t& J, unsigned int size, gi::ex& Tens, MPI_Comm comm = MPI_COMM_WORLD ) {
- unsigned int a1, a2, a3, b1;
+end_code_t multiply_combined_master_initial( tensor3D_t& T, unsigned int size, gi::ex& Tens, MPI_Comm comm = MPI_COMM_WORLD ) {
+ unsigned int a1, a2, a4;
gi::ex A;
gi::lst symbols;
@@ -52,6 +52,7 @@ end_code_t multiply_combined_master_initial( tensor3D_t& T, matrix_int_t& J, uns
j = 0;
int receivedresults = 0;
+ unsigned int N = size/2;
std::vector<parameters_t> input;
std::vector<std::string> results_s;
@@ -59,20 +60,17 @@ end_code_t multiply_combined_master_initial( tensor3D_t& T, matrix_int_t& J, uns
/* Build a list of argument sets */
- for( a1 = 0 ; a1 < size; a1++ ){
+ for( a4 = 0 ; a4 < N ; a4++ ){
i=i+1;
- for( a2 = 0; a2 < size ; a2++ ){
- j=j+1;
- for( a3 = 0 ; a3 < size ; a3++ ){
- A = T[a1][a2][a3];
- for( b1 = 0 ; b1 < size ; b1++ ){
- parameters_t p( A, a1, a2, a3, b1 );
- input.push_back( p );
- }
+ for( a2 = 0; a2 < N ; a2++ ){
+ j=j+1;
+ for( a1 = 0 ; a1 < N ; a1++ ){
+ parameters_t p( a4, a2, a1 );
+ input.push_back( p );
+ }
}
}
- }
-
+
/* Compute the set of symbols */
/* Could be done while the first slave is working */
@@ -90,31 +88,29 @@ end_code_t multiply_combined_master_initial( tensor3D_t& T, matrix_int_t& J, uns
/* Distribute the work */
while( input.size() > 0 ) {
- t_start = rdtsc();
+ t_start = rdtsc();
MPI_Recv( &len, 1, MPI_UNSIGNED, MPI_ANY_SOURCE, MPI_ANY_TAG, comm, &status );
src = status.MPI_SOURCE;
- t_wait = rdtsc() - t_start;
- times[ times_idx] = t_wait;
- times_idx = ( times_idx + 1 ) % ( np - 1 );
- if( !initialround )
- t_average = std::accumulate( times.begin(), times.end(), 0.0 )/(double)(np - 1);
-
- std::cout << "wait " << t_wait << std::endl;
-
+ t_wait = rdtsc() - t_start;
+ times[ times_idx] = t_wait;
+ times_idx = ( times_idx + 1 ) % ( np - 1 );
+ if( !initialround )
+ t_average = std::accumulate( times.begin(), times.end(), 0.0 )/(double)(np - 1);
+
if( status.MPI_TAG == TAG_PULL ) {
/* Nothing else will come: just send wome work */
send_work( input, src, comm );
-
- if( initialround ){
- running++;
- if( np - 1 == running ) initialround = false; // everyone is at work
- }
+
+ if( initialround ){
+ running++;
+ if( np - 1 == running ) initialround = false; // everyone is at work
+ }
} 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 ) {
@@ -125,67 +121,67 @@ end_code_t multiply_combined_master_initial( tensor3D_t& T, matrix_int_t& J, uns
/* 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 );
-
- if( algo == ALGO_MW) {
- send_work( input, src, comm );
+
+ if( algo == ALGO_MW) {
+ send_work( input, src, comm );
}
-
+
/* Process what I have just received */
-
- if( ALGO_MW == algo ) {
- t_start = rdtsc();
- gi::ex received = de_linearize_expression( s, symbols );
- Tens += received;
- t_add = rdtsc() - t_start;
- std::cout << "Add " << t_add << std::endl;
+
+ if( ALGO_MW == algo ) {
+ t_start = rdtsc();
+ gi::ex received = de_linearize_expression( s, symbols );
+ Tens += received;
+ t_add = rdtsc() - t_start;
+ std::cout << "Add " << t_add << std::endl;
#if DEBUG
- results.push_back( received );
- results_s.push_back( s );
- receivedresults++;
+ results.push_back( received );
+ results_s.push_back( s );
+ receivedresults++;
#endif
-
- if( !initialround && t_add > t_average ) {
- /* We are spending too much time adding these results. Now we are going to ask a worker to do this. */
- std::cout << "The master spent too much time computing the sum. Switch to ADDSLAVE algorithm" << std::endl;
- algo = ALGO_ADDSLAVE;
- }
- } else {
- if( ALGO_ADDSLAVE == algo ) {
- results_s.push_back( s );
- send_work_addslave( input, results_s, src );
- } else {
- std::cout << "ERROR: unknown algorithm on the master " << algo << std::endl;
- }
- }
+
+ if( !initialround && t_add > t_average ) {
+ /* We are spending too much time adding these results. Now we are going to ask a worker to do this. */
+ std::cout << "The master spent too much time computing the sum. Switch to ADDSLAVE algorithm" << std::endl;
+ algo = ALGO_ADDSLAVE;
+ }
+ } else {
+ if( ALGO_ADDSLAVE == algo ) {
+ results_s.push_back( s );
+ send_work_addslave( input, results_s, src );
+ } else {
+ std::cout << "ERROR: unknown algorithm on the master " << algo << std::endl;
+ }
+ }
} 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 */
-
+
+ /* 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 );
@@ -222,9 +218,9 @@ end_code_t multiply_combined_master_initial( tensor3D_t& T, matrix_int_t& J, uns
/* The traditional slave */
-void multiply_combined_slave_initial( tensor3D_t& T, matrix_int_t& J, int size, MPI_Comm comm = MPI_COMM_WORLD ) {
+void multiply_combined_slave_initial( tensor3D_t& T, int size, MPI_Comm comm = MPI_COMM_WORLD ) {
gi::ex Tens;
- int a1, a2, a3, b1;
+ int a1, a2, a4;
// gi::ex A;
unsigned int len = 0;
@@ -248,21 +244,19 @@ void multiply_combined_slave_initial( tensor3D_t& T, matrix_int_t& J, int size,
while( true ){
/* Receive a set of parameters */
- t_start = rdtsc();
+ t_start = rdtsc();
MPI_Recv( ¶ms, 1, DT_PARAMETERS, ROOT, MPI_ANY_TAG, comm, &status );
- t_wait = rdtsc() - t_start;
+ t_wait = rdtsc() - t_start;
if( status.MPI_TAG == TAG_WORK ){
a1 = params.a1;
a2 = params.a2;
- a3 = params.a3;
- b1 = params.b1;
- gi::symbol A( std::string( params.A ) );
+ a4 = params.a4;
+
+ t_start = rdtsc();
+ Tens = one_level1_product( &T, size, a4, a2, a1 );
+ t_compute = rdtsc() - t_start;
- t_start = rdtsc();
- Tens = one_level1_product( &T, &J, A, size, a1, a2, a3, b1 );
- t_compute = rdtsc() - t_start;
-
/* TODO if we waited for too long */
if( t_wait > t_compute ) {}
@@ -273,7 +267,7 @@ void multiply_combined_slave_initial( tensor3D_t& T, matrix_int_t& J, int size,
/* Receive a set of expressions to add */
/* Number of expressions received */
- int nb = params.a1;
+ int nb = params.a4;
/* Length of each string */
@@ -291,22 +285,22 @@ void multiply_combined_slave_initial( tensor3D_t& T, matrix_int_t& J, int size,
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 {
+ if( status.MPI_TAG == TAG_END ){
+ return;
+ } else {
std::cerr << "Wrong tag received on slave " << status.MPI_TAG << std::endl;
- }
- }
+ }
+ }
}
}
}
@@ -315,31 +309,31 @@ void multiply_combined_slave_initial( tensor3D_t& T, matrix_int_t& J, int size,
* Combined master-worker *
*******************************************************************************/
-gi::ex multiply_combined_master( tensor3D_t& T, matrix_int_t& J, int size ) { // simpler: same dimension everywhere
+gi::ex multiply_combined_master( tensor3D_t& T, int size ) { // simpler: same dimension everywhere
gi::ex Tens = 0;
end_code_t rc;
/* Initially: start as a traditional M/W */
- rc = multiply_combined_master_initial( T, J, size, Tens );
+ rc = multiply_combined_master_initial( T, size, Tens );
switch( rc ){
case FINISHED:
- return Tens;
+ return Tens;
}
return Tens;
}
-void multiply_combined_worker( tensor3D_t& T, matrix_int_t& J, int size ) { // simpler: same dimension everywhere
+void multiply_combined_worker( tensor3D_t& T, int size ) { // simpler: same dimension everywhere
gi::ex Tens = 0;
std::cout << "worker" << std::endl;
- multiply_combined_slave_initial( T, J, size );
+ multiply_combined_slave_initial( T, size );
}
-gi::ex multiply_combined( tensor3D_t& T, matrix_int_t& J, int size ) { // simpler: same dimension everywhere
+gi::ex multiply_combined( tensor3D_t& T, int size ) { // simpler: same dimension everywhere
int rank;
gi::ex Tens = 0;
@@ -352,9 +346,9 @@ gi::ex multiply_combined( tensor3D_t& T, matrix_int_t& J, int size ) { // simpl
/* Here we go */
if( 0 == rank ) {
- Tens = multiply_combined_master( T, J, size );
+ Tens = multiply_combined_master( T, size );
} else {
- multiply_combined_worker( T, J, size );
+ multiply_combined_worker( T, size );
}
/* Finalize */
diff --git a/src/tensormatrix.h b/src/tensormatrix.h
index 5f471ded12bcb98d474b8a8f59c7af794dc58528..f87ac5ad24e18b14137a99bb34031fc8463d256f 100644
--- a/src/tensormatrix.h
+++ b/src/tensormatrix.h
@@ -26,7 +26,7 @@ gi::ex multiply_2levels( tensor3D_t&, matrix_int_t&, int );
gi::ex multiply_1level_mw( tensor3D_t&, int );
gi::ex multiply_1level_mw_addslave( tensor3D_t&, int );
gi::ex multiply_1level_mw_hierarch( tensor3D_t&, matrix_int_t&, int );
-gi::ex multiply_combined( tensor3D_t&, matrix_int_t&, int );
+gi::ex multiply_combined( tensor3D_t&, int );
/*******************************************************************************
* Default values *
diff --git a/src/tensormatrix_mpi.cpp b/src/tensormatrix_mpi.cpp
index f1aa1dd9581bece858df7c35f480d652b6c196e8..73889e091980cbfe3091804b3f1554c3ab2d2450 100644
--- a/src/tensormatrix_mpi.cpp
+++ b/src/tensormatrix_mpi.cpp
@@ -139,10 +139,10 @@ int main( int argc, char** argv ){
break;
/*case 'h':
Tpara = multiply_1level_mw_hierarch( T, J, N );
- break;
- case 'c':
- Tpara = multiply_combined( T, J, N );
break;*/
+ case 'c':
+ Tpara = multiply_combined( T, N );
+ break;
case 's':
Tpara = multiply_seq( T, J, N );
break;