Venado optimizations (#755)

* Add build script for venado hackathon * Add preliminary method to set matrix size * Add build scripts for hackathon * Add bml_transpose_inplace Fortran subroutine * Modify Fortran bml_transpose API to match the C interface * Modify tests to use new transpose API * Venado build modifications * Move build scripts to scripts/ dir * Expose dense matrix pointer using bml_get_ptr_dense * New introspection methods o Back out bml_get_ptr_dense() from bml_getters o Write fortran wrapper for existing bml_get_data_ptr_dense() method o Write new bml_get_ld_dense() to enable magma matrix pointer use * Remove LANL-specific build scripts * Remove an additional LANL specific build script * Add N_allocated and check in bml_set_N_dense * Reallocate if bml_set_N_dense fails test * Remove LANL-specific build script * Fix linter issues * Fix syntax error in workflow (#758) Signed-off-by: Nicolas Bock <[email protected]> * Update GitHub checkout action (#757) Signed-off-by: Nicolas Bock <[email protected]> * Add yaml extension to dev container (#756) Helpful when editing workflow files. Signed-off-by: Nicolas Bock <[email protected]> * Add another VSCode extension to container (#759) Signed-off-by: Nicolas Bock <[email protected]> * Update version of artifact action (#761) Signed-off-by: Nicolas Bock <[email protected]> * Update compiler for OS X runner (#762) Signed-off-by: Nicolas Bock <[email protected]> * Linter fix --------- Signed-off-by: Nicolas Bock <[email protected]> Co-authored-by: Nicolas Bock <[email protected]>
lanl · Feb 5, 2025 · f386348 · f386348
1 parent bda9a9a
commit f386348
Show file tree

Hide file tree

Showing 12 changed files with 127 additions and 6 deletions.
diff --git a/src/C-interface/dense/bml_allocate_dense_typed.c b/src/C-interface/dense/bml_allocate_dense_typed.c
@@ -98,6 +98,7 @@ bml_matrix_dense_t *TYPED_FUNC(
     A->matrix_type = dense;
     A->matrix_precision = MATRIX_PRECISION;
     A->N = matrix_dimension.N_rows;
+    A->N_allocated = A->N;
     A->distribution_mode = distrib_mode;
 #ifdef BML_USE_MAGMA
     A->ld = magma_roundup(matrix_dimension.N_rows, 32);
@@ -154,6 +155,7 @@ bml_matrix_dense_t *TYPED_FUNC(
     A->matrix_type = dense;
     A->matrix_precision = MATRIX_PRECISION;
     A->N = matrix_dimension.N_rows;
+    A->N_allocated = A->N;
     A->distribution_mode = distrib_mode;
 #ifdef BML_USE_MAGMA
     A->ld = magma_roundup(matrix_dimension.N_rows, 32);

diff --git a/src/C-interface/dense/bml_introspection_dense.c b/src/C-interface/dense/bml_introspection_dense.c
@@ -200,3 +200,22 @@ bml_get_data_ptr_dense(
 {
     return A->matrix;
 }
+
+/** Return the dense matrix ld parameter.
+ *
+ * \param A The dense matrix.
+ * \return The matrix ld parameter.
+ */
+int
+bml_get_ld_dense(
+    bml_matrix_dense_t * A)
+{
+    if (A != NULL)
+    {
+        return A->ld;
+    }
+    else
+    {
+        return -1;
+    }
+}
diff --git a/src/C-interface/dense/bml_setters_dense.c b/src/C-interface/dense/bml_setters_dense.c
@@ -2,6 +2,37 @@
 #include "../bml_logger.h"
 #include "bml_setters_dense.h"
 #include "bml_types_dense.h"
+#include "bml_allocate_dense.h"
+
+#ifdef BML_USE_MAGMA
+#include "magma_v2.h"
+#endif
+
+void
+bml_set_N_dense(
+    bml_matrix_dense_t * A,
+    int N)
+{
+    if (A->N <= A->N_allocated)
+    {
+        A->N = N;
+#ifdef BML_USE_MAGMA
+        A->ld = magma_roundup(A->N, 32);
+#else
+        A->ld = A->N;
+#endif
+    }
+    else
+    {
+        bml_matrix_dense_t *B;
+        bml_matrix_dimension_t matrix_dimension = { A->N, A->N, A->N };
+
+        B = bml_noinit_matrix_dense(A->matrix_precision, matrix_dimension,
+                                    A->distribution_mode);
+        bml_deallocate_dense(A);
+        A = B;
+    }
+}
 
 void
 bml_set_element_dense(

diff --git a/src/C-interface/dense/bml_setters_dense.h b/src/C-interface/dense/bml_setters_dense.h
@@ -7,6 +7,10 @@
 
 #include <complex.h>
 
+void bml_set_N_dense(
+    bml_matrix_dense_t * A,
+    int N);
+
 void bml_set_element_dense(
     bml_matrix_dense_t * A,
     int i,

diff --git a/src/C-interface/dense/bml_types_dense.h b/src/C-interface/dense/bml_types_dense.h
@@ -22,6 +22,8 @@ struct bml_matrix_dense_t
     bml_distribution_mode_t distribution_mode;
     /** The number of rows/columns. */
     int N;
+    /** The number of rows/columns originally allocated. */
+    int N_allocated;
     /** The dense matrix. */
     void *matrix;
     /** The leading dimension of the array matrix. */

diff --git a/src/Fortran-interface/bml_c_interface_m.F90 b/src/Fortran-interface/bml_c_interface_m.F90
@@ -371,6 +371,18 @@ function bml_get_bandwidth_C(a) bind(C, name="bml_get_bandwidth")
       integer(C_INT) :: bml_get_bandwidth_C
     end function bml_get_bandwidth_C
 
+    function bml_get_data_ptr_dense_C(a) bind(C, name="bml_get_data_ptr_dense")
+      import :: C_PTR
+      type(C_PTR), value, intent(in) :: a
+      type(C_PTR) :: bml_get_data_ptr_dense_C
+    end function bml_get_data_ptr_dense_C
+
+    function bml_get_ld_dense_C(a) bind(C, name="bml_get_ld_dense")
+      import :: C_PTR, C_INT
+      type(C_PTR), value, intent(in) :: a
+      integer(C_INT) :: bml_get_ld_dense_C
+    end function bml_get_ld_dense_C
+
     function bml_get_sparsity_C(a, threshold) bind(C, name="bml_get_sparsity")
       import :: C_PTR, C_DOUBLE, C_INT
       type(C_PTR), value, intent(in) :: a
@@ -448,6 +460,12 @@ subroutine bml_scale_inplace_C(alpha, a) bind(C, name="bml_scale_inplace")
       type(C_PTR), value :: a
     end subroutine bml_scale_inplace_C
 
+    subroutine bml_set_N_dense_C(a,n) bind(C, name="bml_set_N_dense")
+      import :: C_PTR, C_INT
+      type(C_PTR), value, intent(in) :: a
+      integer(C_INT), value, intent(in) :: n
+    end subroutine bml_set_N_dense_C
+
     subroutine bml_set_row_C(a, i, row, threshold) bind(C, name="bml_set_row")
       import :: C_PTR, C_INT, C_DOUBLE
       type(C_PTR), value, intent(in) :: a
@@ -616,6 +634,11 @@ function bml_transpose_new_C(a) bind(C, name="bml_transpose_new")
       type(C_PTR) :: bml_transpose_new_C
     end function bml_transpose_new_C
 
+    subroutine bml_transpose_C(a) bind(C, name="bml_transpose")
+      import :: C_PTR
+      type(C_PTR), value, intent(in) :: a
+    end subroutine bml_transpose_C
+
     subroutine bml_write_bml_matrix_C(a, filename) &
          & bind(C, name="bml_write_bml_matrix")
       import :: C_PTR, C_CHAR

diff --git a/src/Fortran-interface/bml_introspection_m.F90 b/src/Fortran-interface/bml_introspection_m.F90
@@ -18,6 +18,8 @@ module bml_introspection_m
   public :: bml_get_bandwidth
   public :: bml_get_distribution_mode
   public :: bml_get_sparsity
+  public :: bml_get_data_ptr_dense
+  public :: bml_get_ld_dense
 
 contains
 
@@ -247,4 +249,25 @@ function bml_get_sparsity(a, threshold) result(sparsity)
 
   end function bml_get_sparsity
 
+  function bml_get_data_ptr_dense(a)
+    type(bml_matrix_t), intent(inout) :: a
+    type(C_PTR) :: bml_get_data_ptr_dense
+
+    bml_get_data_ptr_dense = bml_get_data_ptr_dense_C(a%ptr)
+
+  end function bml_get_data_ptr_dense
+
+  !> Return the dense matrix ld parameter.
+  !!
+  !!\param a The matrix.
+  !!\return The matrix ld parameter.
+  function bml_get_ld_dense(a)
+
+    type(bml_matrix_t), intent(in) :: a
+    integer :: bml_get_ld_dense
+
+    bml_get_ld_dense = bml_get_ld_dense_C(a%ptr)
+
+  end function bml_get_ld_dense
+
 end module bml_introspection_m
diff --git a/src/Fortran-interface/bml_setters_m.F90 b/src/Fortran-interface/bml_setters_m.F90
@@ -52,9 +52,18 @@ module bml_setters_m
   public :: bml_set_diagonal
   public :: bml_set_element
   public :: bml_set_element_new
+  public :: bml_set_N_dense
 
 contains
 
+  subroutine bml_set_N_dense(a,n)
+    type(bml_matrix_t), intent(inout) :: a
+    integer(C_INT), intent(in) :: n
+
+    call bml_set_N_dense_C(a%ptr,n)
+
+  end subroutine bml_set_N_dense
+
   !Setters for element new
   subroutine bml_set_element_new_single_real(a,i,j,element)
 

diff --git a/src/Fortran-interface/bml_transpose_m.F90 b/src/Fortran-interface/bml_transpose_m.F90
@@ -7,22 +7,30 @@ module bml_transpose_m
   implicit none
   private
 
-  public :: bml_transpose
+  public :: bml_transpose_new, bml_transpose
 
 contains
 
   !> Return the transpose of a matrix.
   !!
   !! @param a The matrix.
   !! @param a_t The transpose.
-  subroutine bml_transpose(a, a_t)
+  subroutine bml_transpose_new(a, a_t)
 
     type(bml_matrix_t), intent(in) :: a
     type(bml_matrix_t), intent(inout) :: a_t
 
     call bml_deallocate(a_t)
     a_t%ptr = bml_transpose_new_C(a%ptr)
 
+  end subroutine bml_transpose_new
+
+  subroutine bml_transpose(a)
+
+    type(bml_matrix_t), intent(in) :: a
+
+    call bml_transpose_C(a%ptr)
+
   end subroutine bml_transpose
 
 end module bml_transpose_m
diff --git a/tests/Fortran-tests/diagonalize_matrix_typed.F90 b/tests/Fortran-tests/diagonalize_matrix_typed.F90
@@ -35,7 +35,7 @@ function test_diagonalize_matrix_typed(matrix_type, element_kind, element_precis
     call bml_random_matrix(matrix_type, element_kind, element_precision, n, m, &
          & a)
     call bml_print_matrix("A", a, 1, n, 1, n)
-    call bml_transpose(a, a_t)
+    call bml_transpose_new(a, a_t)
     call bml_print_matrix("A_t", a_t, 1, n, 1, n)
     call bml_add(a, a_t, 0.5d0, 0.5d0, threshold)
     call bml_print_matrix("A", a, 1, n, 1, n)
@@ -44,7 +44,7 @@ function test_diagonalize_matrix_typed(matrix_type, element_kind, element_precis
          & eigenvectors)
     !! \todo Fixme: diagonalization routine is not respecting precision
     call bml_diagonalize(a, eigenvalues, eigenvectors)
-    call bml_transpose(eigenvectors, eigenvectors_t)
+    call bml_transpose_new(eigenvectors, eigenvectors_t)
     call bml_zero_matrix(matrix_type, element_kind, element_precision, n, m, b)
     call bml_zero_matrix(matrix_type, element_kind, element_precision, n, m, c)
     call bml_multiply(eigenvectors_t, eigenvectors, b)

diff --git a/tests/Fortran-tests/io_matrix_typed.F90 b/tests/Fortran-tests/io_matrix_typed.F90
@@ -12,7 +12,7 @@ module io_matrix_typed
 
   public :: test_io_matrix_typed
 
-#ifdef CRAY_SDK
+#if defined(CRAY_SDK) || defined(NVHPC_SDK)
   interface
     integer(c_int) function getpid() bind(c,name="getpid")
       use iso_c_binding

diff --git a/tests/Fortran-tests/transpose_matrix_typed.F90 b/tests/Fortran-tests/transpose_matrix_typed.F90
@@ -35,7 +35,7 @@ function test_transpose_matrix_typed(matrix_type, element_kind, element_precisio
 
     call bml_random_matrix(matrix_type, element_kind, element_precision, n, m, &
          & a)
-    call bml_transpose(a, b)
+    call bml_transpose_new(a, b)
     call bml_copy_new(a, c)
 
     call bml_export_to_dense(a, a_dense)