Add iteration numbers for traces

CMakeLists.txt

@@ -706,6 +706,11 @@ if( CHAMELEON_SCHED_STARPU )
             set(HAVE_STARPU_IDLE_PREFETCH 1)
             message("-- ${Blue}Add definition HAVE_STARPU_IDLE_PREFETCH${ColourReset}")
         endif()
+        check_function_exists(starpu_set_iteration STARPU_SET_ITERATION_FOUND)
+        if ( STARPU_SET_ITERATION_FOUND )
+            set(HAVE_STARPU_SET_ITERATION 1)
+            message("-- ${Blue}Add definition HAVE_STARPU_SET_ITERATION${ColourReset}")
+        endif()
         check_function_exists(starpu_data_wont_use STARPU_DATA_WONT_USE_FOUND)
         if ( STARPU_DATA_WONT_USE_FOUND )
             set(HAVE_STARPU_DATA_WONT_USE 1)

compute/pzgelqf.c

@@ -115,6 +115,8 @@ void morse_pzgelqf(MORSE_desc_t *A, MORSE_desc_t *T,
 #endif
 
     for (k = 0; k < minMNT; k++) {
+        RUNTIME_set_iteration(k);
+
         tempkm = k == A->mt-1 ? A->m-k*A->mb : A->mb;
         tempkn = k == A->nt-1 ? A->n-k*A->nb : A->nb;
         ldak = BLKLDD(A, k);

compute/pzgelqfrh.c

@@ -119,6 +119,8 @@ void morse_pzgelqfrh(MORSE_desc_t *A, MORSE_desc_t *T, int BS,
 #endif
 
     for (k = 0; k < chameleon_min(A->mt, A->nt); k++) {
+        RUNTIME_set_iteration(k);
+
         tempkm = k == A->mt-1 ? A->m-k*A->mb : A->mb;
         ldak = BLKLDD(A, k);
         for (N = k; N < A->nt; N += BS) {

compute/pzgeqrf.c

@@ -110,6 +110,8 @@ void morse_pzgeqrf(MORSE_desc_t *A, MORSE_desc_t *T,
 #endif
 
     for (k = 0; k < minMNT; k++) {
+        RUNTIME_set_iteration(k);
+
         tempkm = k == A->mt-1 ? A->m-k*A->mb : A->mb;
         tempkn = k == A->nt-1 ? A->n-k*A->nb : A->nb;
         ldak = BLKLDD(A, k);

compute/pzgeqrfrh.c

@@ -118,6 +118,8 @@ void morse_pzgeqrfrh(MORSE_desc_t *A, MORSE_desc_t *T, int BS,
 
     K = chameleon_min(A->mt, A->nt);
     for (k = 0; k < K; k++) {
+        RUNTIME_set_iteration(k);
+
         tempkn = k == A->nt-1 ? A->n-k*A->nb : A->nb;
         for (M = k; M < A->mt; M += BS) {
             tempMm = M == A->mt-1 ? A->m-M*A->mb : A->mb;

compute/pzgetrf_incpiv.c

@@ -78,6 +78,8 @@ void morse_pzgetrf_incpiv(MORSE_desc_t *A, MORSE_desc_t *L, int *IPIV,
     morse_zdesc_alloc_diag(*DIAG, A->mb, A->nb, chameleon_min(A->m, A->n), A->nb, 0, 0, chameleon_min(A->m, A->n), A->nb, A->p, A->q);
 
     for (k = 0; k < minMNT; k++) {
+        RUNTIME_set_iteration(k);
+
         tempkm = k == A->mt-1 ? A->m-k*A->mb : A->mb;
         tempkn = k == A->nt-1 ? A->n-k*A->nb : A->nb;
         ldak = BLKLDD(A, k);

compute/pzgetrf_nopiv.c

@@ -54,6 +54,8 @@ void morse_pzgetrf_nopiv(MORSE_desc_t *A,
     ib = MORSE_IB;
 
     for (k = 0; k < chameleon_min(A->mt, A->nt); k++) {
+        RUNTIME_set_iteration(k);
+
         tempkm = k == A->mt-1 ? A->m-k*A->mb : A->mb;
         tempkn = k == A->nt-1 ? A->n-k*A->nb : A->nb;
         ldak = BLKLDD(A, k);

compute/pzhetrd_he2hb.c

@@ -137,6 +137,8 @@ void morse_pzhetrd_he2hb(MORSE_enum uplo,
 
     if (uplo == MorseLower) {
        for (k = 0; k < A->nt-1; k++){
+           RUNTIME_set_iteration(k);
+
            tempkm = k+1 == A->mt-1 ? A->m-(k+1)*A->mb : A->mb;
            tempkn = k   == A->nt-1 ? A->n- k   *A->nb : A->nb;
            ldak1 = BLKLDD(A, k+1);
@@ -290,6 +292,8 @@ void morse_pzhetrd_he2hb(MORSE_enum uplo,
     }
     else {
        for (k = 0; k < A->nt-1; k++){
+           RUNTIME_set_iteration(k);
+
            tempkn = k+1 == A->nt-1 ? A->n-(k+1)*A->nb : A->nb;
            tempkm = k   == A->mt-1 ? A->m- k   *A->mb : A->mb;
            ldak  = BLKLDD(A, k);

compute/pzpotrf.c

@@ -71,6 +71,8 @@ void morse_pzpotrf(MORSE_enum uplo, MORSE_desc_t *A,
      */
     if (uplo == MorseLower) {
         for (k = 0; k < A->mt; k++) {
+            RUNTIME_set_iteration(k);
+
             tempkm = k == A->mt-1 ? A->m-k*A->mb : A->mb;
             ldak = BLKLDD(A, k);
 
@@ -128,6 +130,8 @@ void morse_pzpotrf(MORSE_enum uplo, MORSE_desc_t *A,
      */
     else {
         for (k = 0; k < A->nt; k++) {
+            RUNTIME_set_iteration(k);
+
             tempkm = k == A->nt-1 ? A->n-k*A->nb : A->nb;
             ldak = BLKLDD(A, k);
 

compute/pzpotrimm.c

@@ -76,6 +76,8 @@ void morse_pzpotrimm(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B, MORSE_de
          *  ZPOTRF
          */
         for (k = 0; k < A->mt; k++) {
+            RUNTIME_set_iteration(k);
+
             tempkm = k == A->mt-1 ? A->m-k*A->mb : A->mb;
             ldak = BLKLDD(A, k);
 
@@ -124,6 +126,8 @@ void morse_pzpotrimm(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B, MORSE_de
          *  ZTRTRI
          */
         for (k = 0; k < A->nt; k++) {
+            RUNTIME_set_iteration(A->nt + k);
+
             tempkn = k == A->nt-1 ? A->n-k*A->nb : A->nb;
             ldak = BLKLDD(A, k);
             for (m = k+1; m < A->mt; m++) {
@@ -170,6 +174,8 @@ void morse_pzpotrimm(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B, MORSE_de
          *  ZLAUUM
          */
         for (k = 0; k < A->mt; k++) {
+            RUNTIME_set_iteration(2*A->nt + k);
+
             tempkm = k == A->mt-1 ? A->m-k*A->mb : A->mb;
             ldak = BLKLDD(A, k);
             for(n = 0; n < k; n++) {
@@ -211,6 +217,8 @@ void morse_pzpotrimm(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B, MORSE_de
          *  ZSYMM Right / Lower
          */
         for (k = 0; k < C->nt; k++) {
+            RUNTIME_set_iteration(3*A->nt + k);
+
             tempkn = k == C->nt-1 ? C->n-k*C->nb : C->nb;
             ldak = BLKLDD(A, k);
             zbeta = k == 0 ? beta : zone;
@@ -269,6 +277,8 @@ void morse_pzpotrimm(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B, MORSE_de
          *  ZPOTRF
          */
         for (k = 0; k < A->nt; k++) {
+            RUNTIME_set_iteration(k);
+
             tempkm = k == A->nt-1 ? A->n-k*A->nb : A->nb;
             ldak = BLKLDD(A, k);
             MORSE_TASK_zpotrf(
@@ -317,6 +327,8 @@ void morse_pzpotrimm(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B, MORSE_de
          *  ZTRTRI
          */
         for (k = 0; k < A->mt; k++) {
+            RUNTIME_set_iteration(A->nt + k);
+
             tempkm = k == A->mt-1 ? A->m-k*A->mb : A->mb;
             ldak = BLKLDD(A, k);
             for (n = k+1; n < A->nt; n++) {
@@ -363,6 +375,8 @@ void morse_pzpotrimm(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B, MORSE_de
          *  ZLAUUM
          */
         for (k = 0; k < A->mt; k++) {
+            RUNTIME_set_iteration(2*A->nt + k);
+
             tempkn = k == A->nt-1 ? A->n-k*A->nb : A->nb;
             ldak = BLKLDD(A, k);
 
@@ -406,6 +420,8 @@ void morse_pzpotrimm(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B, MORSE_de
          *  ZSYMM Right / Upper
          */
         for (k = 0; k < C->nt; k++) {
+            RUNTIME_set_iteration(3*A->nt + k);
+
             tempkn = k == C->nt-1 ? C->n-k*C->nb : C->nb;
             ldak = BLKLDD(A, k);
             zbeta = k == 0 ? beta : zone;

compute/pzsytrf.c

@@ -60,6 +60,8 @@ void morse_pzsytrf(MORSE_enum uplo, MORSE_desc_t *A,
      */
     if (uplo == MorseLower) {
         for (k = 0; k < A->mt; k++) {
+            RUNTIME_set_iteration(k);
+
             tempkm = k == A->mt-1 ? A->m-k*A->mb : A->mb;
             ldak = BLKLDD(A, k);
 
@@ -110,6 +112,8 @@ void morse_pzsytrf(MORSE_enum uplo, MORSE_desc_t *A,
      */
     else {
         for (k = 0; k < A->nt; k++) {
+            RUNTIME_set_iteration(k);
+
             tempkm = k == A->nt-1 ? A->n-k*A->nb : A->nb;
             ldak = BLKLDD(A, k);
             MORSE_TASK_zsytrf_nopiv(

compute/pztpgqrt.c

@@ -101,6 +101,8 @@ void morse_pztpgqrt( int L,
 #endif
 
     for (k = V1->nt-1; k >= 0; k--) {
+        RUNTIME_set_iteration(k);
+
         tempkm = k == V1->mt-1 ? V1->m-k*V1->mb : V1->mb;
         tempkk = k == V1->nt-1 ? V1->n-k*V1->nb : V1->nb;
         tempkn = k == Q1->nt-1 ? Q1->n-k*Q1->nb : Q1->nb;

compute/pztpqrt.c

@@ -91,6 +91,8 @@ void morse_pztpqrt( int L, MORSE_desc_t *A, MORSE_desc_t *B, MORSE_desc_t *T,
     RUNTIME_options_ws_alloc( &options, ws_worker, ws_host );
 
     for (k = 0; k < A->nt; k++) {
+        RUNTIME_set_iteration(k);
+
         tempkm = k == A->mt-1 ? A->m-k*A->mb : A->mb;
         tempkn = k == A->nt-1 ? A->n-k*A->nb : A->nb;
         ldak = BLKLDD(A, k);

compute/pztrtri.c

@@ -56,6 +56,8 @@ void morse_pztrtri(MORSE_enum uplo, MORSE_enum diag, MORSE_desc_t *A,
      */
     if (uplo == MorseLower) {
         for (k = 0; k < A->nt; k++) {
+            RUNTIME_set_iteration(k);
+
             tempkn = k == A->nt-1 ? A->n-k*A->nb : A->nb;
             ldak = BLKLDD(A, k);
             for (m = k+1; m < A->mt; m++) {
@@ -104,6 +106,8 @@ void morse_pztrtri(MORSE_enum uplo, MORSE_enum diag, MORSE_desc_t *A,
      */
     else {
         for (k = 0; k < A->mt; k++) {
+            RUNTIME_set_iteration(k);
+
             tempkm = k == A->mt-1 ? A->m-k*A->mb : A->mb;
             ldak = BLKLDD(A, k);
             for (n = k+1; n < A->nt; n++) {

compute/pzunglq.c

@@ -98,6 +98,8 @@ void morse_pzunglq(MORSE_desc_t *A, MORSE_desc_t *Q, MORSE_desc_t *T,
 #endif
 
     for (k = minMT-1; k >= 0; k--) {
+        RUNTIME_set_iteration(k);
+
         tempAkm  = k == A->mt-1 ? A->m-k*A->mb : A->mb;
         tempAkn  = k == A->nt-1 ? A->n-k*A->nb : A->nb;
         tempkmin = chameleon_min( tempAkn, tempAkm );

compute/pzunglqrh.c

@@ -99,6 +99,8 @@ void morse_pzunglqrh(MORSE_desc_t *A, MORSE_desc_t *Q,
 
     K = chameleon_min(A->mt, A->nt);
     for (k = K-1; k >= 0; k--) {
+        RUNTIME_set_iteration(k);
+
         tempkm = k == A->mt-1 ? A->m-k*A->mb : A->mb;
         ldak = BLKLDD(A, k);
         lastRD = 0;

compute/pzungqr.c

@@ -98,6 +98,8 @@ void morse_pzungqr(MORSE_desc_t *A, MORSE_desc_t *Q, MORSE_desc_t *T,
 #endif
 
     for (k = minMT-1; k >= 0; k--) {
+        RUNTIME_set_iteration(k);
+
         tempAkm  = k == A->mt-1 ? A->m-k*A->mb : A->mb;
         tempAkn  = k == A->nt-1 ? A->n-k*A->nb : A->nb;
         tempkmin = chameleon_min( tempAkn, tempAkm );

compute/pzungqrrh.c

@@ -101,6 +101,8 @@ void morse_pzungqrrh(MORSE_desc_t *A, MORSE_desc_t *Q,
 
     K = chameleon_min(A->mt, A->nt);
     for (k = K-1; k >= 0; k--) {
+        RUNTIME_set_iteration(k);
+
         tempkn = k == A->nt-1 ? A->n-k*A->nb : A->nb;
         lastRD = 0;
         for (RD = BS; RD < A->mt-k; RD *= 2)

compute/pzunmlq.c

@@ -105,6 +105,8 @@ void morse_pzunmlq(MORSE_enum side, MORSE_enum trans,
              *  MorseLeft / MorseNoTrans
              */
             for (k = 0; k < minMT; k++) {
+                RUNTIME_set_iteration(k);
+
                 tempkm   = k == B->mt-1 ? B->m-k*B->mb : B->mb;
                 tempkmin = k == minMT-1 ? minM-k*A->nb : A->nb;
                 ldak = BLKLDD(A, k);
@@ -155,6 +157,8 @@ void morse_pzunmlq(MORSE_enum side, MORSE_enum trans,
              *  MorseLeft / MorseConjTrans
              */
             for (k = minMT-1; k >= 0; k--) {
+                RUNTIME_set_iteration(k);
+
                 tempkm   = k == B->mt-1 ? B->m-k*B->mb : B->mb;
                 tempkmin = k == minMT-1 ? minM-k*A->nb : A->nb;
                 ldak = BLKLDD(A, k);
@@ -207,6 +211,8 @@ void morse_pzunmlq(MORSE_enum side, MORSE_enum trans,
              *  MorseRight / MorseNoTrans
              */
             for (k = minMT-1; k >= 0; k--) {
+                RUNTIME_set_iteration(k);
+
                 tempkn   = k == B->nt -1 ? B->n -k*B->nb : B->nb;
                 tempkmin = k == minMT-1 ? minM-k*A->nb : A->nb;
                 ldak = BLKLDD(A, k);
@@ -257,6 +263,8 @@ void morse_pzunmlq(MORSE_enum side, MORSE_enum trans,
              *  MorseRight / MorseConjTrans
              */
             for (k = 0; k < minMT; k++) {
+                RUNTIME_set_iteration(k);
+
                 tempkn   = k == B->nt -1 ? B->n -k*B->nb : B->nb;
                 tempkmin = k == minMT-1 ? minM-k*A->mb : A->mb;
                 ldak = BLKLDD(A, k);

compute/pzunmlqrh.c

@@ -105,6 +105,8 @@ void morse_pzunmlqrh(MORSE_enum side, MORSE_enum trans,
              *  MorseLeft / MorseNoTrans
              */
             for (k = 0; k < K; k++) {
+                RUNTIME_set_iteration(k);
+
                 tempkm = k == A->mt-1 ? A->m-k*A->mb : A->mb;
                 ldak = BLKLDD(A, k);
                 for (N = k; N < A->nt; N += BS) {
@@ -178,6 +180,8 @@ void morse_pzunmlqrh(MORSE_enum side, MORSE_enum trans,
              *  MorseLeft / MorseConjTrans
              */
             for (k = K-1; k >= 0; k--) {
+                RUNTIME_set_iteration(k);
+
                 tempkm = k == A->mt-1 ? A->m-k*A->mb : A->mb;
                 ldak = BLKLDD(A, k);
                 lastRD = 0;
@@ -257,6 +261,8 @@ void morse_pzunmlqrh(MORSE_enum side, MORSE_enum trans,
              *  MorseRight / MorseNoTrans
              */
             for (k = K-1; k >= 0; k--) {
+                RUNTIME_set_iteration(k);
+
                 tempkm = k == A->mt-1 ? A->m-k*A->mb : A->mb;
                 ldak = BLKLDD(A, k);
                 lastRD = 0;
@@ -332,6 +338,8 @@ void morse_pzunmlqrh(MORSE_enum side, MORSE_enum trans,
              *  MorseRight / MorseConjTrans
              */
             for (k = 0; k < K; k++) {
+                RUNTIME_set_iteration(k);
+
                 tempkm = k == A->mt-1 ? A->m-k*A->mb : A->mb;
                 ldak = BLKLDD(A, k);
                 for (N = k; N < A->nt; N += BS) {

compute/pzunmqr.c

@@ -105,6 +105,8 @@ void morse_pzunmqr(MORSE_enum side, MORSE_enum trans,
              *  MorseLeft / MorseConjTrans
              */
             for (k = 0; k < minMT; k++) {
+                RUNTIME_set_iteration(k);
+
                 tempkm   = k == B->mt-1 ? B->m-k*B->mb : B->mb;
                 tempkmin = k == minMT-1 ? minM-k*A->nb : A->nb;
                 ldak = BLKLDD(A, k);
@@ -156,6 +158,8 @@ void morse_pzunmqr(MORSE_enum side, MORSE_enum trans,
          */
         else {
             for (k = minMT-1; k >= 0; k--) {
+                RUNTIME_set_iteration(k);
+
                 tempkm = k == B->mt-1 ? B->m-k*B->mb : B->mb;
                 tempkmin = k == minMT-1 ? minM-k*A->nb : A->nb;
                 ldak = BLKLDD(A, k);
@@ -209,6 +213,8 @@ void morse_pzunmqr(MORSE_enum side, MORSE_enum trans,
     else {
         if (trans == MorseConjTrans) {
             for (k = minMT-1; k >= 0; k--) {
+                RUNTIME_set_iteration(k);
+
                 tempkn = k == B->nt-1 ? B->n-k*B->nb : B->nb;
                 tempkmin = k == minMT-1 ? minM-k*A->nb : A->nb;