return inverse permutatuin for gpund and gpumnd

owensgroup · Sep 16, 2024 · 95bcdc7 · 95bcdc7
1 parent 41cee4d
commit 95bcdc7
Show file tree

Hide file tree

Showing 4 changed files with 43 additions and 20 deletions.
diff --git a/apps/NDReorder/test_all_permutations.cu b/apps/NDReorder/test_all_permutations.cu
@@ -38,18 +38,6 @@ void no_permute(rxmesh::RXMeshStatic& rx, const EigeMatT& eigen_mat)
     RXMESH_INFO(" No-permutation NNZ = {}", nnz);
 }
 
-template <typename T>
-std::vector<T> inverse_permutation(const std::vector<T>& perm)
-{
-    std::vector<int> perm_inv(perm.size());
-
-    for (int i = 0; i < perm_inv.size(); ++i) {
-        perm_inv[perm[i]] = i;
-    }
-
-    return perm_inv;
-}
-
 
 template <typename T, typename EigeMatT>
 void with_metis(rxmesh::RXMeshStatic&          rx,
@@ -160,6 +148,10 @@ void with_gpumgnd(rxmesh::RXMeshStatic& rx, const EigeMatT& eigen_mat)
     EXPECT_TRUE(
         rxmesh::is_unique_permutation(h_permute.size(), h_permute.data()));
 
+    std::vector<int> helper(rx.get_num_vertices());
+    rxmesh::inverse_permutation(
+        rx.get_num_vertices(), h_permute.data(), helper.data());
+
     // render_permutation(rx, h_permute, "GPUMGND");
 
     int nnz = count_nnz_fillin(eigen_mat, h_permute, "gpumgnd");
@@ -177,6 +169,10 @@ void with_gpu_nd(rxmesh::RXMeshStatic& rx, const EigeMatT& eigen_mat)
     EXPECT_TRUE(
         rxmesh::is_unique_permutation(h_permute.size(), h_permute.data()));
 
+    std::vector<int> helper(rx.get_num_vertices());
+    rxmesh::inverse_permutation(
+        rx.get_num_vertices(), h_permute.data(), helper.data());
+
     // render_permutation(rx, h_permute, "GPUND");
 
     int nnz = count_nnz_fillin(eigen_mat, h_permute, "gpund");
@@ -201,15 +197,17 @@ void with_amd(rxmesh::RXMeshStatic&    rx,
                 h_permute.size() * sizeof(int));
 
 
-    std::vector<int> h_permute_inv = inverse_permutation(h_permute);
+    std::vector<int> helper(rx.get_num_vertices());
+    rxmesh::inverse_permutation(
+        rx.get_num_vertices(), h_permute.data(), helper.data());
 
 
-    EXPECT_TRUE(rxmesh::is_unique_permutation(h_permute_inv.size(),
-                                              h_permute_inv.data()));
+    EXPECT_TRUE(
+        rxmesh::is_unique_permutation(h_permute.size(), h_permute.data()));
 
-    // render_permutation(rx, h_permute_inv, "AMD");
+    // render_permutation(rx, h_permute, "AMD");
 
-    int nnz = count_nnz_fillin(eigen_mat, h_permute_inv, "amd");
+    int nnz = count_nnz_fillin(eigen_mat, h_permute, "amd");
 
     RXMESH_INFO(" With AMD NNZ = {}", nnz);
 }
@@ -233,11 +231,13 @@ void with_symrcm(rxmesh::RXMeshStatic&    rx,
     EXPECT_TRUE(
         rxmesh::is_unique_permutation(h_permute.size(), h_permute.data()));
 
-    std::vector<int> h_permute_inv = inverse_permutation(h_permute);
+    std::vector<int> helper(rx.get_num_vertices());
+    rxmesh::inverse_permutation(
+        rx.get_num_vertices(), h_permute.data(), helper.data());
 
-    // render_permutation(rx, h_permute_inv, "symrcm");
+    // render_permutation(rx, h_permute, "symrcm");
 
-    int nnz = count_nnz_fillin(eigen_mat, h_permute_inv, "symrcm");
+    int nnz = count_nnz_fillin(eigen_mat, h_permute, "symrcm");
 
     RXMESH_INFO(" With SYMRCM NNZ = {}", nnz);
 }

diff --git a/include/rxmesh/matrix/mgnd_permute.cuh b/include/rxmesh/matrix/mgnd_permute.cuh
@@ -7,6 +7,8 @@
 #include <thrust/execution_policy.h>
 #include <thrust/scan.h>
 
+#include "rxmesh/matrix/permute_util.h"
+
 namespace rxmesh {
 
 namespace detail {
@@ -150,6 +152,9 @@ inline void mgnd_permute(const RXMeshStatic& rx, int* h_permute)
                           rx.get_num_vertices() * sizeof(int),
                           cudaMemcpyDeviceToHost));
 
+    std::vector<int> helper(rx.get_num_vertices());
+    inverse_permutation(rx.get_num_vertices(), h_permute, helper.data());
+
     GPU_FREE(d_v_ordering_prefix_sum);
     GPU_FREE(d_permute);
 }

diff --git a/include/rxmesh/matrix/nd_permute.cuh b/include/rxmesh/matrix/nd_permute.cuh
@@ -940,6 +940,10 @@ void nd_permute(RXMeshStatic& rx, int* h_permute)
                           rx.get_num_vertices() * sizeof(int),
                           cudaMemcpyDeviceToHost));
 
+
+    std::vector<int> helper(rx.get_num_vertices());
+    inverse_permutation(rx.get_num_vertices(), h_permute, helper.data());
+
     GPU_FREE(d_permute);
     GPU_FREE(d_patch_proj_l);
     GPU_FREE(d_patch_proj_l1);

diff --git a/include/rxmesh/matrix/permute_util.h b/include/rxmesh/matrix/permute_util.h
@@ -26,4 +26,18 @@ bool is_unique_permutation(uint32_t size, T* h_permute)
 
     return true;
 }
+
+/**
+ * @brief given a permutation array, overwrite it with its inverse
+ */
+template <typename T>
+void inverse_permutation(uint32_t size, T* perm, T* helper)
+{
+    for (int i = 0; i < size; ++i) {
+        helper[perm[i]] = i;
+    }
+    for (int i = 0; i < size; ++i) {
+        perm[i] = helper[i];
+    }
+}
 }  // namespace rxmesh