void cube_blas_d_iamax (cube_t *ctx, int n, const double *x, int incx, int *result) { cublasStatus_t status; if (! cube_context_check (ctx)) return; status = cublasIdamax (ctx->h_blas, n, x, incx, result); cube_blas_check (ctx, status); }
int main( int argc, char** argv ) { TESTING_INIT(); real_Double_t gflops, t1, t2; double c_neg_one = MAGMA_D_NEG_ONE; magma_int_t ione = 1; const char trans[] = { 'N', 'C', 'T' }; const char uplo[] = { 'L', 'U' }; const char diag[] = { 'U', 'N' }; const char side[] = { 'L', 'R' }; double *A, *B, *C, *C2, *LU; double *dA, *dB, *dC1, *dC2; double alpha = MAGMA_D_MAKE( 0.5, 0.1 ); double beta = MAGMA_D_MAKE( 0.7, 0.2 ); double dalpha = 0.6; double dbeta = 0.8; double work[1], error, total_error; magma_int_t ISEED[4] = {0,0,0,1}; magma_int_t m, n, k, size, maxn, ld, info; magma_int_t *piv; magma_err_t err; magma_opts opts; parse_opts( argc, argv, &opts ); printf( "Compares magma wrapper function to cublas function; all diffs should be exactly 0.\n\n" ); total_error = 0.; for( int i = 0; i < opts.ntest; ++i ) { m = opts.msize[i]; n = opts.nsize[i]; k = opts.ksize[i]; printf("=========================================================================\n"); printf( "M %d, N %d, K %d\n", (int) m, (int) n, (int) k ); // allocate matrices // over-allocate so they can be any combination of {m,n,k} x {m,n,k}. maxn = max( max( m, n ), k ); ld = maxn; size = maxn*maxn; err = magma_malloc_cpu( (void**) &piv, maxn*sizeof(magma_int_t) ); assert( err == 0 ); err = magma_dmalloc_pinned( &A, size ); assert( err == 0 ); err = magma_dmalloc_pinned( &B, size ); assert( err == 0 ); err = magma_dmalloc_pinned( &C, size ); assert( err == 0 ); err = magma_dmalloc_pinned( &C2, size ); assert( err == 0 ); err = magma_dmalloc_pinned( &LU, size ); assert( err == 0 ); err = magma_dmalloc( &dA, size ); assert( err == 0 ); err = magma_dmalloc( &dB, size ); assert( err == 0 ); err = magma_dmalloc( &dC1, size ); assert( err == 0 ); err = magma_dmalloc( &dC2, size ); assert( err == 0 ); // initialize matrices size = maxn*maxn; lapackf77_dlarnv( &ione, ISEED, &size, A ); lapackf77_dlarnv( &ione, ISEED, &size, B ); lapackf77_dlarnv( &ione, ISEED, &size, C ); printf( "========== Level 1 BLAS ==========\n" ); // ----- test DSWAP // swap 2nd and 3rd columns of dA, then copy to C2 and compare with A assert( n >= 4 ); magma_dsetmatrix( m, n, A, ld, dA, ld ); magma_dsetmatrix( m, n, A, ld, dB, ld ); magma_dswap( m, dA(0,1), 1, dA(0,2), 1 ); magma_dswap( m, dB(0,1), 1, dB(0,2), 1 ); // check results, storing diff between magma and cuda calls in C2 cublasDaxpy( ld*n, c_neg_one, dA, 1, dB, 1 ); magma_dgetmatrix( m, n, dB, ld, C2, ld ); error = lapackf77_dlange( "F", &m, &k, C2, &ld, work ); total_error += error; printf( "dswap diff %.2g\n", error ); // ----- test IDAMAX // get argmax of column of A magma_dsetmatrix( m, k, A, ld, dA, ld ); error = 0; for( int j = 0; j < k; ++j ) { magma_int_t i1 = magma_idamax( m, dA(0,j), 1 ); magma_int_t i2 = cublasIdamax( m, dA(0,j), 1 ); assert( i1 == i2 ); error += abs( i1 - i2 ); } total_error += error; gflops = (double)m * k / 1e9; printf( "idamax diff %.2g\n", error ); printf( "\n" ); printf( "========== Level 2 BLAS ==========\n" ); // ----- test DGEMV // c = alpha*A*b + beta*c, with A m*n; b,c m or n-vectors // try no-trans/trans for( int ia = 0; ia < 3; ++ia ) { magma_dsetmatrix( m, n, A, ld, dA, ld ); magma_dsetvector( maxn, B, 1, dB, 1 ); magma_dsetvector( maxn, C, 1, dC1, 1 ); magma_dsetvector( maxn, C, 1, dC2, 1 ); t1 = magma_sync_wtime( 0 ); magma_dgemv( trans[ia], m, n, alpha, dA, ld, dB, 1, beta, dC1, 1 ); t1 = magma_sync_wtime( 0 ) - t1; t2 = magma_sync_wtime( 0 ); cublasDgemv( trans[ia], m, n, alpha, dA, ld, dB, 1, beta, dC2, 1 ); t2 = magma_sync_wtime( 0 ) - t2; // check results, storing diff between magma and cuda call in C2 size = (trans[ia] == 'N' ? m : n); cublasDaxpy( size, c_neg_one, dC1, 1, dC2, 1 ); magma_dgetvector( size, dC2, 1, C2, 1 ); error = lapackf77_dlange( "F", &size, &ione, C2, &ld, work ); total_error += error; gflops = FLOPS_DGEMV( m, n ) / 1e9; printf( "dgemv( %c ) diff %.2g, Gflop/s %6.2f, %6.2f\n", trans[ia], error, gflops/t1, gflops/t2 ); } printf( "\n" ); // ----- test DSYMV // c = alpha*A*b + beta*c, with A m*m symmetric; b,c m-vectors // try upper/lower for( int iu = 0; iu < 2; ++iu ) { magma_dsetmatrix( m, m, A, ld, dA, ld ); magma_dsetvector( m, B, 1, dB, 1 ); magma_dsetvector( m, C, 1, dC1, 1 ); magma_dsetvector( m, C, 1, dC2, 1 ); t1 = magma_sync_wtime( 0 ); magma_dsymv( uplo[iu], m, alpha, dA, ld, dB, 1, beta, dC1, 1 ); t1 = magma_sync_wtime( 0 ) - t1; t2 = magma_sync_wtime( 0 ); cublasDsymv( uplo[iu], m, alpha, dA, ld, dB, 1, beta, dC2, 1 ); t2 = magma_sync_wtime( 0 ) - t2; // check results, storing diff between magma and cuda call in C2 cublasDaxpy( m, c_neg_one, dC1, 1, dC2, 1 ); magma_dgetvector( m, dC2, 1, C2, 1 ); error = lapackf77_dlange( "F", &m, &ione, C2, &ld, work ); total_error += error; gflops = FLOPS_DSYMV( m ) / 1e9; printf( "dsymv( %c ) diff %.2g, Gflop/s %6.2f, %6.2f\n", uplo[iu], error, gflops/t1, gflops/t2 ); } printf( "\n" ); // ----- test DTRSV // solve A*c = c, with A m*m triangular; c m-vector // try upper/lower, no-trans/trans, unit/non-unit diag // Factor A into LU to get well-conditioned triangles, else solve yields garbage. // Still can give garbage if solves aren't consistent with LU factors, // e.g., using unit diag for U, so copy lower triangle to upper triangle. // Also used for trsm later. lapackf77_dlacpy( "Full", &maxn, &maxn, A, &ld, LU, &ld ); lapackf77_dgetrf( &maxn, &maxn, LU, &ld, piv, &info ); for( int j = 0; j < maxn; ++j ) { for( int i = 0; i < j; ++i ) { *LU(i,j) = *LU(j,i); } } for( int iu = 0; iu < 2; ++iu ) { for( int it = 0; it < 3; ++it ) { for( int id = 0; id < 2; ++id ) { magma_dsetmatrix( m, m, LU, ld, dA, ld ); magma_dsetvector( m, C, 1, dC1, 1 ); magma_dsetvector( m, C, 1, dC2, 1 ); t1 = magma_sync_wtime( 0 ); magma_dtrsv( uplo[iu], trans[it], diag[id], m, dA, ld, dC1, 1 ); t1 = magma_sync_wtime( 0 ) - t1; t2 = magma_sync_wtime( 0 ); cublasDtrsv( uplo[iu], trans[it], diag[id], m, dA, ld, dC2, 1 ); t2 = magma_sync_wtime( 0 ) - t2; // check results, storing diff between magma and cuda call in C2 cublasDaxpy( m, c_neg_one, dC1, 1, dC2, 1 ); magma_dgetvector( m, dC2, 1, C2, 1 ); error = lapackf77_dlange( "F", &m, &ione, C2, &ld, work ); total_error += error; gflops = FLOPS_DTRSM( MagmaLeft, m, 1 ) / 1e9; printf( "dtrsv( %c, %c, %c ) diff %.2g, Gflop/s %6.2f, %6.2f\n", uplo[iu], trans[it], diag[id], error, gflops/t1, gflops/t2 ); }}} printf( "\n" ); printf( "========== Level 3 BLAS ==========\n" ); // ----- test DGEMM // C = alpha*A*B + beta*C, with A m*k or k*m; B k*n or n*k; C m*n // try combinations of no-trans/trans for( int ia = 0; ia < 3; ++ia ) { for( int ib = 0; ib < 3; ++ib ) { bool nta = (trans[ia] == 'N'); bool ntb = (trans[ib] == 'N'); magma_dsetmatrix( (nta ? m : k), (nta ? m : k), A, ld, dA, ld ); magma_dsetmatrix( (ntb ? k : n), (ntb ? n : k), B, ld, dB, ld ); magma_dsetmatrix( m, n, C, ld, dC1, ld ); magma_dsetmatrix( m, n, C, ld, dC2, ld ); t1 = magma_sync_wtime( 0 ); magma_dgemm( trans[ia], trans[ib], m, n, k, alpha, dA, ld, dB, ld, beta, dC1, ld ); t1 = magma_sync_wtime( 0 ) - t1; t2 = magma_sync_wtime( 0 ); cublasDgemm( trans[ia], trans[ib], m, n, k, alpha, dA, ld, dB, ld, beta, dC2, ld ); t2 = magma_sync_wtime( 0 ) - t2; // check results, storing diff between magma and cuda call in C2 cublasDaxpy( ld*n, c_neg_one, dC1, 1, dC2, 1 ); magma_dgetmatrix( m, n, dC2, ld, C2, ld ); error = lapackf77_dlange( "F", &m, &n, C2, &ld, work ); total_error += error; gflops = FLOPS_DGEMM( m, n, k ) / 1e9; printf( "dgemm( %c, %c ) diff %.2g, Gflop/s %6.2f, %6.2f\n", trans[ia], trans[ib], error, gflops/t1, gflops/t2 ); }} printf( "\n" ); // ----- test DSYMM // C = alpha*A*B + beta*C (left) with A m*m symmetric; B,C m*n; or // C = alpha*B*A + beta*C (right) with A n*n symmetric; B,C m*n // try left/right, upper/lower for( int is = 0; is < 2; ++is ) { for( int iu = 0; iu < 2; ++iu ) { magma_dsetmatrix( m, m, A, ld, dA, ld ); magma_dsetmatrix( m, n, B, ld, dB, ld ); magma_dsetmatrix( m, n, C, ld, dC1, ld ); magma_dsetmatrix( m, n, C, ld, dC2, ld ); t1 = magma_sync_wtime( 0 ); magma_dsymm( side[is], uplo[iu], m, n, alpha, dA, ld, dB, ld, beta, dC1, ld ); t1 = magma_sync_wtime( 0 ) - t1; t2 = magma_sync_wtime( 0 ); cublasDsymm( side[is], uplo[iu], m, n, alpha, dA, ld, dB, ld, beta, dC2, ld ); t2 = magma_sync_wtime( 0 ) - t2; // check results, storing diff between magma and cuda call in C2 cublasDaxpy( ld*n, c_neg_one, dC1, 1, dC2, 1 ); magma_dgetmatrix( m, n, dC2, ld, C2, ld ); error = lapackf77_dlange( "F", &m, &n, C2, &ld, work ); total_error += error; gflops = FLOPS_DSYMM( side[is], m, n ) / 1e9; printf( "dsymm( %c, %c ) diff %.2g, Gflop/s %6.2f, %6.2f\n", side[is], uplo[iu], error, gflops/t1, gflops/t2 ); }} printf( "\n" ); // ----- test DSYRK // C = alpha*A*A^H + beta*C (no-trans) with A m*k and C m*m symmetric; or // C = alpha*A^H*A + beta*C (trans) with A k*m and C m*m symmetric // try upper/lower, no-trans/trans for( int iu = 0; iu < 2; ++iu ) { for( int it = 0; it < 3; ++it ) { magma_dsetmatrix( n, k, A, ld, dA, ld ); magma_dsetmatrix( n, n, C, ld, dC1, ld ); magma_dsetmatrix( n, n, C, ld, dC2, ld ); t1 = magma_sync_wtime( 0 ); magma_dsyrk( uplo[iu], trans[it], n, k, dalpha, dA, ld, dbeta, dC1, ld ); t1 = magma_sync_wtime( 0 ) - t1; t2 = magma_sync_wtime( 0 ); cublasDsyrk( uplo[iu], trans[it], n, k, dalpha, dA, ld, dbeta, dC2, ld ); t2 = magma_sync_wtime( 0 ) - t2; // check results, storing diff between magma and cuda call in C2 cublasDaxpy( ld*n, c_neg_one, dC1, 1, dC2, 1 ); magma_dgetmatrix( n, n, dC2, ld, C2, ld ); error = lapackf77_dlange( "F", &n, &n, C2, &ld, work ); total_error += error; gflops = FLOPS_DSYRK( k, n ) / 1e9; printf( "dsyrk( %c, %c ) diff %.2g, Gflop/s %6.2f, %6.2f\n", uplo[iu], trans[it], error, gflops/t1, gflops/t2 ); }} printf( "\n" ); // ----- test DSYR2K // C = alpha*A*B^H + ^alpha*B*A^H + beta*C (no-trans) with A,B n*k; C n*n symmetric; or // C = alpha*A^H*B + ^alpha*B^H*A + beta*C (trans) with A,B k*n; C n*n symmetric // try upper/lower, no-trans/trans for( int iu = 0; iu < 2; ++iu ) { for( int it = 0; it < 3; ++it ) { bool nt = (trans[it] == 'N'); magma_dsetmatrix( (nt ? n : k), (nt ? n : k), A, ld, dA, ld ); magma_dsetmatrix( n, n, C, ld, dC1, ld ); magma_dsetmatrix( n, n, C, ld, dC2, ld ); t1 = magma_sync_wtime( 0 ); magma_dsyr2k( uplo[iu], trans[it], n, k, alpha, dA, ld, dB, ld, dbeta, dC1, ld ); t1 = magma_sync_wtime( 0 ) - t1; t2 = magma_sync_wtime( 0 ); cublasDsyr2k( uplo[iu], trans[it], n, k, alpha, dA, ld, dB, ld, dbeta, dC2, ld ); t2 = magma_sync_wtime( 0 ) - t2; // check results, storing diff between magma and cuda call in C2 cublasDaxpy( ld*n, c_neg_one, dC1, 1, dC2, 1 ); magma_dgetmatrix( n, n, dC2, ld, C2, ld ); error = lapackf77_dlange( "F", &n, &n, C2, &ld, work ); total_error += error; gflops = FLOPS_DSYR2K( k, n ) / 1e9; printf( "dsyr2k( %c, %c ) diff %.2g, Gflop/s %6.2f, %6.2f\n", uplo[iu], trans[it], error, gflops/t1, gflops/t2 ); }} printf( "\n" ); // ----- test DTRMM // C = alpha*A*C (left) with A m*m triangular; C m*n; or // C = alpha*C*A (right) with A n*n triangular; C m*n // try left/right, upper/lower, no-trans/trans, unit/non-unit for( int is = 0; is < 2; ++is ) { for( int iu = 0; iu < 2; ++iu ) { for( int it = 0; it < 3; ++it ) { for( int id = 0; id < 2; ++id ) { bool left = (side[is] == 'L'); magma_dsetmatrix( (left ? m : n), (left ? m : n), A, ld, dA, ld ); magma_dsetmatrix( m, n, C, ld, dC1, ld ); magma_dsetmatrix( m, n, C, ld, dC2, ld ); t1 = magma_sync_wtime( 0 ); magma_dtrmm( side[is], uplo[iu], trans[it], diag[id], m, n, alpha, dA, ld, dC1, ld ); t1 = magma_sync_wtime( 0 ) - t1; t2 = magma_sync_wtime( 0 ); cublasDtrmm( side[is], uplo[iu], trans[it], diag[id], m, n, alpha, dA, ld, dC2, ld ); t2 = magma_sync_wtime( 0 ) - t2; // check results, storing diff between magma and cuda call in C2 cublasDaxpy( ld*n, c_neg_one, dC1, 1, dC2, 1 ); magma_dgetmatrix( m, n, dC2, ld, C2, ld ); error = lapackf77_dlange( "F", &n, &n, C2, &ld, work ); total_error += error; gflops = FLOPS_DTRMM( side[is], m, n ) / 1e9; printf( "dtrmm( %c, %c ) diff %.2g, Gflop/s %6.2f, %6.2f\n", uplo[iu], trans[it], error, gflops/t1, gflops/t2 ); }}}} printf( "\n" ); // ----- test DTRSM // solve A*X = alpha*B (left) with A m*m triangular; B m*n; or // solve X*A = alpha*B (right) with A n*n triangular; B m*n // try left/right, upper/lower, no-trans/trans, unit/non-unit for( int is = 0; is < 2; ++is ) { for( int iu = 0; iu < 2; ++iu ) { for( int it = 0; it < 3; ++it ) { for( int id = 0; id < 2; ++id ) { bool left = (side[is] == 'L'); magma_dsetmatrix( (left ? m : n), (left ? m : n), LU, ld, dA, ld ); magma_dsetmatrix( m, n, C, ld, dC1, ld ); magma_dsetmatrix( m, n, C, ld, dC2, ld ); t1 = magma_sync_wtime( 0 ); magma_dtrsm( side[is], uplo[iu], trans[it], diag[id], m, n, alpha, dA, ld, dC1, ld ); t1 = magma_sync_wtime( 0 ) - t1; t2 = magma_sync_wtime( 0 ); cublasDtrsm( side[is], uplo[iu], trans[it], diag[id], m, n, alpha, dA, ld, dC2, ld ); t2 = magma_sync_wtime( 0 ) - t2; // check results, storing diff between magma and cuda call in C2 cublasDaxpy( ld*n, c_neg_one, dC1, 1, dC2, 1 ); magma_dgetmatrix( m, n, dC2, ld, C2, ld ); error = lapackf77_dlange( "F", &n, &n, C2, &ld, work ); total_error += error; gflops = FLOPS_DTRSM( side[is], m, n ) / 1e9; printf( "dtrsm( %c, %c ) diff %.2g, Gflop/s %6.2f, %6.2f\n", uplo[iu], trans[it], error, gflops/t1, gflops/t2 ); }}}} printf( "\n" ); // cleanup magma_free_cpu( piv ); magma_free_pinned( A ); magma_free_pinned( B ); magma_free_pinned( C ); magma_free_pinned( C2 ); magma_free_pinned( LU ); magma_free( dA ); magma_free( dB ); magma_free( dC1 ); magma_free( dC2 ); } if ( total_error != 0. ) { printf( "total error %.2g -- ought to be 0 -- some test failed (see above).\n", total_error ); } else { printf( "all tests passed\n" ); } TESTING_FINALIZE(); return 0; }
int iamax(const Vector<double> &x) { return cublasIdamax(x.getSize(), x, x.inc()); }
magma_int_t magma_idamax( magma_int_t n, const double *dx, magma_int_t incx ) { return cublasIdamax( n, dx, incx ); }
// // Overloaded function for dispatching to // * CUBLAS backend, and // * double value-type. // inline std::ptrdiff_t iamax( const int n, const double* x, const int incx ) { return cublasIdamax( n, x, incx ); }
int CORE_dtstrf_cublas(int M, int N, int IB, int NB, double *U, int LDU, double *A, int LDA, double *L, int LDL, int *IPIV, double *WORK, int LDWORK, int *INFO) { static double zzero = 0.0; static double mzone =-1.0; cublasStatus_t status; cudaError_t err; double alpha; int i, j, ii, sb; int im, ip; #if CONFIG_VERBOSE fprintf(stdout, "%s: M=%d N=%d IB=%d NB=%d U=%p LDU=%d A=%p LDA=%d L=%p LDL=%d IPIV=%p WORK=%p LDWORK=%d\n", __FUNCTION__, M, N, IB, NB, U, LDU, A, LDA, L, LDL, IPIV, WORK, LDWORK); fflush(stdout); #endif /* Check input arguments */ *INFO = 0; if (M < 0) { coreblas_error(1, "Illegal value of M"); return -1; } if (N < 0) { coreblas_error(2, "Illegal value of N"); return -2; } if (IB < 0) { coreblas_error(3, "Illegal value of IB"); return -3; } if ((LDU < max(1,NB)) && (NB > 0)) { coreblas_error(6, "Illegal value of LDU"); return -6; } if ((LDA < max(1,M)) && (M > 0)) { coreblas_error(8, "Illegal value of LDA"); return -8; } if ((LDL < max(1,IB)) && (IB > 0)) { coreblas_error(10, "Illegal value of LDL"); return -10; } /* Quick return */ if ((M == 0) || (N == 0) || (IB == 0)) return PLASMA_SUCCESS; /* Set L to 0 */ err = cudaMemset(L, 0, LDL*N*sizeof(double)); PLASMA_CUDA_ASSERT(err); double* dev_ptr = 0; err = cudaMalloc((void**)&dev_ptr, 2*sizeof(double)); PLASMA_CUDA_ASSERT(err); double* host_ptr; err = cudaMallocHost((void**)&host_ptr, 2*sizeof(double)); PLASMA_CUDA_ASSERT(err); int* piv = kaapi_memory_get_host_pointer_and_validate(IPIV); ip = 0; for (ii = 0; ii < N; ii += IB) { sb = min(N-ii, IB); for (i = 0; i < sb; i++) { status = cublasIdamax(kaapi_cuda_cublas_handle(), M, &A[LDA*(ii+i)], 1, &im ); PLASMA_CUBLAS_ASSERT(status); /* get im */ err = cudaStreamSynchronize(kaapi_cuda_kernel_stream()); PLASMA_CUDA_ASSERT(err); /* ajust index, CUBLAS is 1-based indexing */ im--; piv[ip] = ii+i+1; core_dtstrf_cmp(kaapi_cuda_kernel_stream(), &A[LDA*(ii+i)+im], &U[LDU*(ii+i)+ii+i], dev_ptr, host_ptr); err = cudaStreamSynchronize(kaapi_cuda_kernel_stream()); PLASMA_CUDA_ASSERT(err); if (host_ptr[0] == 1.0f) { /* * Swap behind. */ status = cublasDswap(kaapi_cuda_cublas_handle(), i, &L[LDL*ii+i], LDL, &WORK[im], LDWORK ); PLASMA_CUBLAS_ASSERT(status); /* * Swap ahead. */ status = cublasDswap(kaapi_cuda_cublas_handle(), sb-i, &U[LDU*(ii+i)+ii+i], LDU, &A[LDA*(ii+i)+im], LDA ); PLASMA_CUBLAS_ASSERT(status); /* * Set IPIV. */ piv[ip] = NB + im + 1; core_dtstrf_set_zero(kaapi_cuda_kernel_stream(), A, LDA, i, ii, im, zzero ); } core_dtstrf_cmp_zzero_and_get_alpha(kaapi_cuda_kernel_stream(), &A[LDA*(ii+i)+im], &U[LDU*(ii+i)+ii+i], zzero, dev_ptr, host_ptr); err = cudaStreamSynchronize(kaapi_cuda_kernel_stream()); PLASMA_CUDA_ASSERT(err); if ((*INFO == 0) && (host_ptr[0] == 1.0f)) { *INFO = ii+i+1; } // alpha = ((double)1. / U[LDU*(ii+i)+ii+i]); alpha = host_ptr[1]; status = cublasDscal(kaapi_cuda_cublas_handle(), M, &alpha, &A[LDA*(ii+i)], 1 ); PLASMA_CUBLAS_ASSERT(status); status = cublasDcopy(kaapi_cuda_cublas_handle(), M, &A[LDA*(ii+i)], 1, &WORK[LDWORK*i], 1 ); PLASMA_CUBLAS_ASSERT(status); status = cublasDger(kaapi_cuda_cublas_handle(), M, sb-i-1, &mzone, &A[LDA*(ii+i)], 1, &U[LDU*(ii+i+1)+ii+i], LDU, &A[LDA*(ii+i+1)], LDA ); PLASMA_CUBLAS_ASSERT(status); ip = ip+1; } /* * Apply the subpanel to the rest of the panel. */ if(ii+i < N) { for(j = ii; j < ii+sb; j++) { if (piv[j] <= NB) { piv[j] = piv[j] - ii; } } CORE_dssssm_cublas_v2( NB, N-(ii+sb), M, N-(ii+sb), sb, sb, &U[LDU*(ii+sb)+ii], LDU, &A[LDA*(ii+sb)], LDA, &L[LDL*ii], LDL, WORK, LDWORK, &piv[ii] ); err = cudaStreamSynchronize(kaapi_cuda_kernel_stream()); PLASMA_CUDA_ASSERT(err); for(j = ii; j < ii+sb; j++) { if (piv[j] <= NB) { piv[j] = piv[j] + ii; } } } } cudaFreeHost(host_ptr); cudaFree(dev_ptr); return PLASMA_SUCCESS; }