Skip to content

Commit e0d0528

Browse files
joseemoreirajoseemoreira
committed
conditional debug messages
1 parent 132ada4 commit e0d0528

File tree

1 file changed

+37
-34
lines changed

1 file changed

+37
-34
lines changed

‎Code/OptionG/portable.cc‎

Lines changed: 37 additions & 34 deletions
Original file line numberDiff line numberDiff line change
@@ -576,36 +576,36 @@ void microdgemm
576576
double *B0 = B; double *B1 = B0 + RV->VLENE(); double *B2 = B1 + RV->VLENE(); double *B3 = B2 + RV->VLENE();
577577
for (u32 k=0; k<K; k+=lambda_eff)
578578
{
579-
std::cout << "k = " << k << std::endl;
579+
if (debug > 1) { std::cout << "k = " << k << std::endl; }
580580

581-
vmtlfre64.v( 0, A0, lambda_eff); { std::cout << "VR[ 0] = "; RV->printVRf64( 0); }
582-
vmtlfre64.v( 1, A1, lambda_eff); { std::cout << "VR[ 1] = "; RV->printVRf64( 1); }
583-
vmtlfre64.v( 2, A2, lambda_eff); { std::cout << "VR[ 2] = "; RV->printVRf64( 2); }
584-
vmtlfre64.v( 3, A3, lambda_eff); { std::cout << "VR[ 3] = "; RV->printVRf64( 3); }
585-
vmtlfre64.v( 8, B0, lambda_eff); { std::cout << "VR[ 8] = "; RV->printVRf64( 8); }
586-
vmtlfre64.v( 9, B1, lambda_eff); { std::cout << "VR[ 9] = "; RV->printVRf64( 9); }
587-
vmtlfre64.v(10, B2, lambda_eff); { std::cout << "VR[10] = "; RV->printVRf64(10); }
588-
vmtlfre64.v(11, B3, lambda_eff); { std::cout << "VR[11] = "; RV->printVRf64(11); }
581+
vmtlfre64.v( 0, A0, lambda_eff); if (debug > 1) { std::cout << "VR[ 0] = "; RV->printVRf64( 0); }
582+
vmtlfre64.v( 1, A1, lambda_eff); if (debug > 1) { std::cout << "VR[ 1] = "; RV->printVRf64( 1); }
583+
vmtlfre64.v( 2, A2, lambda_eff); if (debug > 1) { std::cout << "VR[ 2] = "; RV->printVRf64( 2); }
584+
vmtlfre64.v( 3, A3, lambda_eff); if (debug > 1) { std::cout << "VR[ 3] = "; RV->printVRf64( 3); }
585+
vmtlfre64.v( 8, B0, lambda_eff); if (debug > 1) { std::cout << "VR[ 8] = "; RV->printVRf64( 8); }
586+
vmtlfre64.v( 9, B1, lambda_eff); if (debug > 1) { std::cout << "VR[ 9] = "; RV->printVRf64( 9); }
587+
vmtlfre64.v(10, B2, lambda_eff); if (debug > 1) { std::cout << "VR[10] = "; RV->printVRf64(10); }
588+
vmtlfre64.v(11, B3, lambda_eff); if (debug > 1) { std::cout << "VR[11] = "; RV->printVRf64(11); }
589589

590590
A0 = A0 + INCA ; A1 = A1 + INCA ; A2 = A2 + INCA ; A3 = A3 + INCA;
591591
B0 = B0 + INCB ; B1 = B1 + INCB ; B2 = B2 + INCB ; B3 = B3 + INCB;
592592

593-
vfmmacc.v0(16, 0, 8); vmrotate.vv( 8, 8); { std::cout << "VR[16] = "; RV->printVRf64(16); }
594-
vfmmacc.v0(17, 0, 9); vmrotate.vv( 9, 9); { std::cout << "VR[17] = "; RV->printVRf64(17); }
595-
vfmmacc.v0(18, 1, 8); vmrotate.vv( 8, 8); { std::cout << "VR[18] = "; RV->printVRf64(18); }
596-
vfmmacc.v0(19, 1, 9); vmrotate.vv( 9, 9); { std::cout << "VR[19] = "; RV->printVRf64(19); }
597-
vfmmacc.v0(20, 0, 10); vmrotate.vv(10, 10); { std::cout << "VR[20] = "; RV->printVRf64(20); }
598-
vfmmacc.v0(21, 0, 11); vmrotate.vv(11, 11); { std::cout << "VR[21] = "; RV->printVRf64(21); }
599-
vfmmacc.v0(22, 1, 10); vmrotate.vv(10, 10); { std::cout << "VR[22] = "; RV->printVRf64(22); }
600-
vfmmacc.v0(23, 1, 11); vmrotate.vv(11, 11); { std::cout << "VR[23] = "; RV->printVRf64(23); }
601-
vfmmacc.v0(24, 2, 8); vmrotate.vv( 8, 8); { std::cout << "VR[24] = "; RV->printVRf64(24); }
602-
vfmmacc.v0(25, 2, 9); vmrotate.vv( 9, 9); { std::cout << "VR[25] = "; RV->printVRf64(25); }
603-
vfmmacc.v0(26, 3, 8); vmrotate.vv( 8, 8); { std::cout << "VR[26] = "; RV->printVRf64(26); }
604-
vfmmacc.v0(27, 3, 9); vmrotate.vv( 9, 9); { std::cout << "VR[27] = "; RV->printVRf64(27); }
605-
vfmmacc.v0(28, 2, 10); vmrotate.vv(10, 10); { std::cout << "VR[28] = "; RV->printVRf64(28); }
606-
vfmmacc.v0(29, 2, 11); vmrotate.vv(11, 11); { std::cout << "VR[29] = "; RV->printVRf64(29); }
607-
vfmmacc.v0(30, 3, 10); vmrotate.vv(10, 10); { std::cout << "VR[30] = "; RV->printVRf64(30); }
608-
vfmmacc.v0(31, 3, 11); vmrotate.vv(11, 11); { std::cout << "VR[31] = "; RV->printVRf64(31); }
593+
vfmmacc.v0(16, 0, 8); vmrotate.vv( 8, 8); if (debug > 1) { std::cout << "VR[16] = "; RV->printVRf64(16); }
594+
vfmmacc.v0(17, 0, 9); vmrotate.vv( 9, 9); if (debug > 1) { std::cout << "VR[17] = "; RV->printVRf64(17); }
595+
vfmmacc.v0(18, 1, 8); vmrotate.vv( 8, 8); if (debug > 1) { std::cout << "VR[18] = "; RV->printVRf64(18); }
596+
vfmmacc.v0(19, 1, 9); vmrotate.vv( 9, 9); if (debug > 1) { std::cout << "VR[19] = "; RV->printVRf64(19); }
597+
vfmmacc.v0(20, 0, 10); vmrotate.vv(10, 10); if (debug > 1) { std::cout << "VR[20] = "; RV->printVRf64(20); }
598+
vfmmacc.v0(21, 0, 11); vmrotate.vv(11, 11); if (debug > 1) { std::cout << "VR[21] = "; RV->printVRf64(21); }
599+
vfmmacc.v0(22, 1, 10); vmrotate.vv(10, 10); if (debug > 1) { std::cout << "VR[22] = "; RV->printVRf64(22); }
600+
vfmmacc.v0(23, 1, 11); vmrotate.vv(11, 11); if (debug > 1) { std::cout << "VR[23] = "; RV->printVRf64(23); }
601+
vfmmacc.v0(24, 2, 8); vmrotate.vv( 8, 8); if (debug > 1) { std::cout << "VR[24] = "; RV->printVRf64(24); }
602+
vfmmacc.v0(25, 2, 9); vmrotate.vv( 9, 9); if (debug > 1) { std::cout << "VR[25] = "; RV->printVRf64(25); }
603+
vfmmacc.v0(26, 3, 8); vmrotate.vv( 8, 8); if (debug > 1) { std::cout << "VR[26] = "; RV->printVRf64(26); }
604+
vfmmacc.v0(27, 3, 9); vmrotate.vv( 9, 9); if (debug > 1) { std::cout << "VR[27] = "; RV->printVRf64(27); }
605+
vfmmacc.v0(28, 2, 10); vmrotate.vv(10, 10); if (debug > 1) { std::cout << "VR[28] = "; RV->printVRf64(28); }
606+
vfmmacc.v0(29, 2, 11); vmrotate.vv(11, 11); if (debug > 1) { std::cout << "VR[29] = "; RV->printVRf64(29); }
607+
vfmmacc.v0(30, 3, 10); vmrotate.vv(10, 10); if (debug > 1) { std::cout << "VR[30] = "; RV->printVRf64(30); }
608+
vfmmacc.v0(31, 3, 11); vmrotate.vv(11, 11); if (debug > 1) { std::cout << "VR[31] = "; RV->printVRf64(31); }
609609
}
610610

611611
u32 offset[32];
@@ -626,7 +626,7 @@ void microdgemm
626626
offset[30] = (1 != lmul) ? (offset[29] + RV->lambda()) : (LisSquare() ? offset[28] + RV->lambda() * gamma : offset[28] + RV->lambda());
627627
offset[31] = ((!LisSquare()) && (1 == lmul)) ? offset[29] + RV->lambda() : offset[30] + RV->lambda();
628628

629-
for (u32 i=16; i<32; i++) std::cout << "offset[" << i << "] = " << offset[i] << std::endl;
629+
if (debug > 1) { for (u32 i=16; i<32; i++) std::cout << "offset[" << i << "] = " << offset[i] << std::endl; }
630630

631631
double *D = new double[M*N];
632632
vsetvl(5, 0, 64, 1, true, true); // double-precision kernel, set VL to VLENE and LMUL to 1
@@ -655,8 +655,8 @@ void microdgemm
655655
}
656656
for (u32 i=0; i<M; i++) for (u32 j=0; j<N; j++) C[i*gamma + j] += alpha * S[i*N + j];
657657

658-
std::cout << "S = "; print(M, N, S);
659-
std::cout << "D = "; print(M, N, D);
658+
if (debug > 1) { std::cout << "S = "; print(M, N, S); }
659+
if (debug > 1) { std::cout << "D = "; print(M, N, D); }
660660

661661
for (u32 i=0; i<M; i++) for (u32 j=0; j<N; j++) if (D[i*N+j] != S[i*N+j]) { std::cout << "Error for D[" << i << "," << j << "] = " << D[i*N+j] << " != " << S[i*N+j] << std::endl; exit(-1); }
662662
for (u32 i=0; i<M; i++) for (u32 j=0; j<N; j++) if (E[i*N+j] != C[i*N+j]) { std::cout << "Error for E[" << i << "," << j << "] = " << E[i*N+j] << " != " << C[i*N+j] << std::endl; exit(-1); }
@@ -757,16 +757,16 @@ bool run_microgemm
757757
u32 M = mu;
758758
u32 N = nu;
759759

760-
double alpha = 3.141592654; alpha = 1.0;
760+
double alpha = 3.141592654; // alpha = 1.0;
761761

762762
// Allocate A, B, and C panels
763763
double *A = new double[M*K]; for (u32 i=0; i<M*K; i++) A[i] = i; // drand48() - 0.5;
764764
double *B = new double[K*N]; for (u32 i=0; i<K*N; i++) B[i] = i; // drand48() - 0.5;
765765
double *C = new double[M*N]; for (u32 i=0; i<M*N; i++) C[i] = 0; // drand48() - 0.5;
766766
double *D = new double[M*N]; for (u32 i=0; i<M*N; i++) D[i] = C[i];
767767

768-
std::cout << "A = "; print(M, K, A);
769-
std::cout << "B = "; print(N, K, B);
768+
if (debug > 1) { std::cout << "A = "; print(M, K, A); }
769+
if (debug > 1) { std::cout << "B = "; print(N, K, B); }
770770

771771
// Allocate the packed panels
772772
double *Ap = new double[M*K];
@@ -780,8 +780,8 @@ bool run_microgemm
780780

781781
for (u32 k=0; k<K; k+=lambda_eff)
782782
{
783-
std::cout << "Ap[" << k/lambda_eff << "] = "; print(M, lambda_eff, Ap+k*mu);
784-
std::cout << "Bp[" << k/lambda_eff << "] = "; print(N, lambda_eff, Bp+k*nu);
783+
if (debug > 1) { std::cout << "Ap[" << k/lambda_eff << "] = "; print(M, lambda_eff, Ap+k*mu); }
784+
if (debug > 1) { std::cout << "Bp[" << k/lambda_eff << "] = "; print(N, lambda_eff, Bp+k*nu); }
785785
}
786786

787787
microdgemm(M, N, K, Ap, Bp, alpha, D, N, LMUL);
@@ -856,9 +856,12 @@ int main
856856
run_microgemm<1024, 4>(4);
857857
run_microgemm<1024, 4>(8);
858858
run_microgemm<1024, 4>(16);
859+
run_microgemm<2048, 2>(8);
860+
run_microgemm<2048, 2>(16);
859861
run_microgemm<2048, 4>(8);
860862
run_microgemm<2048, 4>(16);
861-
run_microgemm<4096, 4>(8);
863+
run_microgemm<4096, 2>(16);
864+
run_microgemm<4096, 4>(16);
862865
run_microgemm<4096, 8>(16);
863866

864867
return 0;

0 commit comments

Comments
 (0)