diff --git a/kleidicv/include/kleidicv/filter_driver_neon.h b/kleidicv/include/kleidicv/filter_driver_neon.h
index daacde7afb313e5f8ab106260d05ce0677e7981d..42c857dfd1c39c0e780554944ae886c5ce0dd565 100644
--- a/kleidicv/include/kleidicv/filter_driver_neon.h
+++ b/kleidicv/include/kleidicv/filter_driver_neon.h
@@ -37,6 +37,26 @@ class SeparableFilterDriver {
                                          SourceVecTraits::num_lanes()};
     constexpr auto seq = std::make_index_sequence<KernelSize>{};
 
+    if constexpr (KernelSize == 3) {
+      loop.unroll_twice([&](size_t index) {
+        auto src_0_x2 = vld1q_x2(&src_rows.at(border_offsets.c(0))[index]);
+        auto src_1_x2 = vld1q_x2(&src_rows.at(border_offsets.c(1))[index]);
+        auto src_2_x2 = vld1q_x2(&src_rows.at(border_offsets.c(2))[index]);
+
+        SourceVectorType src_a[3], src_b[3];
+        src_a[0] = src_0_x2.val[0];
+        src_b[0] = src_0_x2.val[1];
+        src_a[1] = src_1_x2.val[0];
+        src_b[1] = src_1_x2.val[1];
+        src_a[2] = src_2_x2.val[0];
+        src_b[2] = src_2_x2.val[1];
+
+        filter_.vertical_vector_path(src_a, &dst_rows[index]);
+        filter_.vertical_vector_path(
+            src_b, &dst_rows[index + SourceVecTraits::num_lanes()]);
+      });
+    }
+
     loop.unroll_once([&](size_t index) {
       vertical_vector_path(src_rows, dst_rows, border_offsets, index, seq);
     });
@@ -54,7 +74,7 @@ class SeparableFilterDriver {
     constexpr auto seq = std::make_index_sequence<KernelSize>{};
 
     loop.unroll_twice([&](size_t index) {
-      horizontal_vector_path_2x(src_rows, dst_rows, border_offsets, index, seq);
+      horizontal_vector_path_2x(src_rows, dst_rows, border_offsets, index);
     });
 
     loop.unroll_once([&](size_t index) {
@@ -97,19 +117,129 @@ class SeparableFilterDriver {
     filter_.vertical_scalar_path(src, &dst_rows[index]);
   }
 
-  template <size_t... SeqNum>
   void horizontal_vector_path_2x(Rows<const BufferType> src_rows,
                                  Rows<DestinationType> dst_rows,
-                                 BorderOffsets border_offsets, size_t index,
-                                 std::index_sequence<SeqNum...>) const {
-    BufferVectorType src_a[KernelSize] = {
-        vld1q(&src_rows.at(0, border_offsets.c(SeqNum))[index])...};
-    BufferVectorType src_b[KernelSize] = {vld1q(&src_rows.at(
-        0, border_offsets.c(SeqNum))[index + BufferVecTraits::num_lanes()])...};
+                                 BorderOffsets border_offsets,
+                                 size_t index) const {
+    if constexpr (KernelSize == 3) {
+      auto src_0_x2 = vld1q_x2(&src_rows.at(0, border_offsets.c(0))[index]);
+      auto src_1_x2 = vld1q_x2(&src_rows.at(0, border_offsets.c(1))[index]);
+      auto src_2_x2 = vld1q_x2(&src_rows.at(0, border_offsets.c(2))[index]);
+
+      BufferVectorType src_a[3], src_b[3];
+      src_a[0] = src_0_x2.val[0];
+      src_b[0] = src_0_x2.val[1];
+      src_a[1] = src_1_x2.val[0];
+      src_b[1] = src_1_x2.val[1];
+      src_a[2] = src_2_x2.val[0];
+      src_b[2] = src_2_x2.val[1];
+
+      filter_.horizontal_vector_path(src_a, &dst_rows[index]);
+      filter_.horizontal_vector_path(
+          src_b, &dst_rows[index + BufferVecTraits::num_lanes()]);
+    } else if constexpr (KernelSize == 5) {
+      BufferVectorType src_a[5], src_b[5];
+      src_a[0] = vld1q(&src_rows.at(0, border_offsets.c(0))[index]);
+      src_b[0] = vld1q(&src_rows.at(
+          0, border_offsets.c(0))[index + BufferVecTraits::num_lanes()]);
+      src_a[1] = vld1q(&src_rows.at(0, border_offsets.c(1))[index]);
+      src_b[1] = vld1q(&src_rows.at(
+          0, border_offsets.c(1))[index + BufferVecTraits::num_lanes()]);
+      src_a[2] = vld1q(&src_rows.at(0, border_offsets.c(2))[index]);
+      src_b[2] = vld1q(&src_rows.at(
+          0, border_offsets.c(2))[index + BufferVecTraits::num_lanes()]);
+      src_a[3] = vld1q(&src_rows.at(0, border_offsets.c(3))[index]);
+      src_b[3] = vld1q(&src_rows.at(
+          0, border_offsets.c(3))[index + BufferVecTraits::num_lanes()]);
+      src_a[4] = vld1q(&src_rows.at(0, border_offsets.c(4))[index]);
+      src_b[4] = vld1q(&src_rows.at(
+          0, border_offsets.c(4))[index + BufferVecTraits::num_lanes()]);
 
-    filter_.horizontal_vector_path(src_a, &dst_rows[index]);
-    filter_.horizontal_vector_path(
-        src_b, &dst_rows[index + BufferVecTraits::num_lanes()]);
+      filter_.horizontal_vector_path(src_a, &dst_rows[index]);
+      filter_.horizontal_vector_path(
+          src_b, &dst_rows[index + BufferVecTraits::num_lanes()]);
+    } else if constexpr (KernelSize == 7) {
+      BufferVectorType src_a[7], src_b[7];
+      src_a[0] = vld1q(&src_rows.at(0, border_offsets.c(0))[index]);
+      src_b[0] = vld1q(&src_rows.at(
+          0, border_offsets.c(0))[index + BufferVecTraits::num_lanes()]);
+      src_a[1] = vld1q(&src_rows.at(0, border_offsets.c(1))[index]);
+      src_b[1] = vld1q(&src_rows.at(
+          0, border_offsets.c(1))[index + BufferVecTraits::num_lanes()]);
+      src_a[2] = vld1q(&src_rows.at(0, border_offsets.c(2))[index]);
+      src_b[2] = vld1q(&src_rows.at(
+          0, border_offsets.c(2))[index + BufferVecTraits::num_lanes()]);
+      src_a[3] = vld1q(&src_rows.at(0, border_offsets.c(3))[index]);
+      src_b[3] = vld1q(&src_rows.at(
+          0, border_offsets.c(3))[index + BufferVecTraits::num_lanes()]);
+      src_a[4] = vld1q(&src_rows.at(0, border_offsets.c(4))[index]);
+      src_b[4] = vld1q(&src_rows.at(
+          0, border_offsets.c(4))[index + BufferVecTraits::num_lanes()]);
+      src_a[5] = vld1q(&src_rows.at(0, border_offsets.c(5))[index]);
+      src_b[5] = vld1q(&src_rows.at(
+          0, border_offsets.c(5))[index + BufferVecTraits::num_lanes()]);
+      src_a[6] = vld1q(&src_rows.at(0, border_offsets.c(6))[index]);
+      src_b[6] = vld1q(&src_rows.at(
+          0, border_offsets.c(6))[index + BufferVecTraits::num_lanes()]);
+
+      filter_.horizontal_vector_path(src_a, &dst_rows[index]);
+      filter_.horizontal_vector_path(
+          src_b, &dst_rows[index + BufferVecTraits::num_lanes()]);
+    } else if constexpr (KernelSize == 15) {
+      BufferVectorType src_a[15], src_b[15];
+      src_a[0] = vld1q(&src_rows.at(0, border_offsets.c(0))[index]);
+      src_b[0] = vld1q(&src_rows.at(
+          0, border_offsets.c(0))[index + BufferVecTraits::num_lanes()]);
+      src_a[1] = vld1q(&src_rows.at(0, border_offsets.c(1))[index]);
+      src_b[1] = vld1q(&src_rows.at(
+          0, border_offsets.c(1))[index + BufferVecTraits::num_lanes()]);
+      src_a[2] = vld1q(&src_rows.at(0, border_offsets.c(2))[index]);
+      src_b[2] = vld1q(&src_rows.at(
+          0, border_offsets.c(2))[index + BufferVecTraits::num_lanes()]);
+      src_a[3] = vld1q(&src_rows.at(0, border_offsets.c(3))[index]);
+      src_b[3] = vld1q(&src_rows.at(
+          0, border_offsets.c(3))[index + BufferVecTraits::num_lanes()]);
+      src_a[4] = vld1q(&src_rows.at(0, border_offsets.c(4))[index]);
+      src_b[4] = vld1q(&src_rows.at(
+          0, border_offsets.c(4))[index + BufferVecTraits::num_lanes()]);
+      src_a[5] = vld1q(&src_rows.at(0, border_offsets.c(5))[index]);
+      src_b[5] = vld1q(&src_rows.at(
+          0, border_offsets.c(5))[index + BufferVecTraits::num_lanes()]);
+      src_a[6] = vld1q(&src_rows.at(0, border_offsets.c(6))[index]);
+      src_b[6] = vld1q(&src_rows.at(
+          0, border_offsets.c(6))[index + BufferVecTraits::num_lanes()]);
+      src_a[7] = vld1q(&src_rows.at(0, border_offsets.c(7))[index]);
+      src_b[7] = vld1q(&src_rows.at(
+          0, border_offsets.c(7))[index + BufferVecTraits::num_lanes()]);
+      src_a[8] = vld1q(&src_rows.at(0, border_offsets.c(8))[index]);
+      src_b[8] = vld1q(&src_rows.at(
+          0, border_offsets.c(8))[index + BufferVecTraits::num_lanes()]);
+      src_a[9] = vld1q(&src_rows.at(0, border_offsets.c(9))[index]);
+      src_b[9] = vld1q(&src_rows.at(
+          0, border_offsets.c(9))[index + BufferVecTraits::num_lanes()]);
+      src_a[10] = vld1q(&src_rows.at(0, border_offsets.c(10))[index]);
+      src_b[10] = vld1q(&src_rows.at(
+          0, border_offsets.c(10))[index + BufferVecTraits::num_lanes()]);
+      src_a[11] = vld1q(&src_rows.at(0, border_offsets.c(11))[index]);
+      src_b[11] = vld1q(&src_rows.at(
+          0, border_offsets.c(11))[index + BufferVecTraits::num_lanes()]);
+      src_a[12] = vld1q(&src_rows.at(0, border_offsets.c(12))[index]);
+      src_b[12] = vld1q(&src_rows.at(
+          0, border_offsets.c(12))[index + BufferVecTraits::num_lanes()]);
+      src_a[13] = vld1q(&src_rows.at(0, border_offsets.c(13))[index]);
+      src_b[13] = vld1q(&src_rows.at(
+          0, border_offsets.c(13))[index + BufferVecTraits::num_lanes()]);
+      src_a[14] = vld1q(&src_rows.at(0, border_offsets.c(14))[index]);
+      src_b[14] = vld1q(&src_rows.at(
+          0, border_offsets.c(14))[index + BufferVecTraits::num_lanes()]);
+
+      filter_.horizontal_vector_path(src_a, &dst_rows[index]);
+      filter_.horizontal_vector_path(
+          src_b, &dst_rows[index + BufferVecTraits::num_lanes()]);
+    } else {
+      static_assert(KernelSize != KernelSize,
+                    "please define variants for other kernel sizes");
+    }
   }
 
   template <size_t... SeqNum>
diff --git a/kleidicv/include/kleidicv/filter_driver_sc.h b/kleidicv/include/kleidicv/filter_driver_sc.h
index 6bbb96eee513d7b58e2f97bf93037e6285512f3d..4333b0c5cc6e92b94100698d1a6c09182d576854 100644
--- a/kleidicv/include/kleidicv/filter_driver_sc.h
+++ b/kleidicv/include/kleidicv/filter_driver_sc.h
@@ -65,7 +65,7 @@ class SeparableFilterDriver {
 
     loop.unroll_twice([&](size_t index) KLEIDICV_STREAMING_COMPATIBLE {
       horizontal_vector_path_2x(pg_all, src_rows, dst_rows, border_offsets,
-                                index, seq);
+                                index);
     });
 
     loop.unroll_once([&](size_t index) KLEIDICV_STREAMING_COMPATIBLE {
@@ -103,20 +103,140 @@ class SeparableFilterDriver {
         &dst_rows[index]);
   }
 
-  template <size_t... SeqNum>
-  void horizontal_vector_path_2x(svbool_t pg, Rows<const BufferType> src_rows,
-                                 Rows<DestinationType> dst_rows,
-                                 BorderOffsets border_offsets, size_t index,
-                                 std::index_sequence<SeqNum...>) const
-      KLEIDICV_STREAMING_COMPATIBLE {
-    filter_.horizontal_vector_path(
-        pg, svld1(pg, &src_rows.at(0, border_offsets.c(SeqNum))[index])...,
-        &dst_rows[index]);
+  void horizontal_vector_path_2x(
+      svbool_t pg, Rows<const BufferType> src_rows,
+      Rows<DestinationType> dst_rows, BorderOffsets border_offsets,
+      size_t index) const KLEIDICV_STREAMING_COMPATIBLE {
+    if constexpr (KernelSize == 3) {
+      auto src_0 = &src_rows.at(0, border_offsets.c(0))[index];
+      auto src_1 = &src_rows.at(0, border_offsets.c(1))[index];
+      auto src_2 = &src_rows.at(0, border_offsets.c(2))[index];
 
-    filter_.horizontal_vector_path(
-        pg,
-        svld1_vnum(pg, &src_rows.at(0, border_offsets.c(SeqNum))[index], 1)...,
-        &dst_rows[index + BufferVecTraits::num_lanes()]);
+      BufferVectorType src_0_0 = svld1(pg, &src_0[0]);
+      BufferVectorType src_1_0 = svld1_vnum(pg, &src_0[0], 1);
+      BufferVectorType src_0_1 = svld1(pg, &src_1[0]);
+      BufferVectorType src_1_1 = svld1_vnum(pg, &src_1[0], 1);
+      BufferVectorType src_0_2 = svld1(pg, &src_2[0]);
+      BufferVectorType src_1_2 = svld1_vnum(pg, &src_2[0], 1);
+
+      filter_.horizontal_vector_path(pg, src_0_0, src_0_1, src_0_2,
+                                     &dst_rows[index]);
+      filter_.horizontal_vector_path(
+          pg, src_1_0, src_1_1, src_1_2,
+          &dst_rows[index + BufferVecTraits::num_lanes()]);
+    } else if constexpr (KernelSize == 5) {
+      auto src_0 = &src_rows.at(0, border_offsets.c(0))[index];
+      auto src_1 = &src_rows.at(0, border_offsets.c(1))[index];
+      auto src_2 = &src_rows.at(0, border_offsets.c(2))[index];
+      auto src_3 = &src_rows.at(0, border_offsets.c(3))[index];
+      auto src_4 = &src_rows.at(0, border_offsets.c(4))[index];
+
+      BufferVectorType src_0_0 = svld1(pg, &src_0[0]);
+      BufferVectorType src_1_0 = svld1_vnum(pg, &src_0[0], 1);
+      BufferVectorType src_0_1 = svld1(pg, &src_1[0]);
+      BufferVectorType src_1_1 = svld1_vnum(pg, &src_1[0], 1);
+      BufferVectorType src_0_2 = svld1(pg, &src_2[0]);
+      BufferVectorType src_1_2 = svld1_vnum(pg, &src_2[0], 1);
+      BufferVectorType src_0_3 = svld1(pg, &src_3[0]);
+      BufferVectorType src_1_3 = svld1_vnum(pg, &src_3[0], 1);
+      BufferVectorType src_0_4 = svld1(pg, &src_4[0]);
+      BufferVectorType src_1_4 = svld1_vnum(pg, &src_4[0], 1);
+
+      filter_.horizontal_vector_path(pg, src_0_0, src_0_1, src_0_2, src_0_3,
+                                     src_0_4, &dst_rows[index]);
+      filter_.horizontal_vector_path(
+          pg, src_1_0, src_1_1, src_1_2, src_1_3, src_1_4,
+          &dst_rows[index + BufferVecTraits::num_lanes()]);
+    } else if constexpr (KernelSize == 7) {
+      auto src_0 = &src_rows.at(0, border_offsets.c(0))[index];
+      auto src_1 = &src_rows.at(0, border_offsets.c(1))[index];
+      auto src_2 = &src_rows.at(0, border_offsets.c(2))[index];
+      auto src_3 = &src_rows.at(0, border_offsets.c(3))[index];
+      auto src_4 = &src_rows.at(0, border_offsets.c(4))[index];
+      auto src_5 = &src_rows.at(0, border_offsets.c(5))[index];
+      auto src_6 = &src_rows.at(0, border_offsets.c(6))[index];
+
+      BufferVectorType src_0_0 = svld1(pg, &src_0[0]);
+      BufferVectorType src_1_0 = svld1_vnum(pg, &src_0[0], 1);
+      BufferVectorType src_0_1 = svld1(pg, &src_1[0]);
+      BufferVectorType src_1_1 = svld1_vnum(pg, &src_1[0], 1);
+      BufferVectorType src_0_2 = svld1(pg, &src_2[0]);
+      BufferVectorType src_1_2 = svld1_vnum(pg, &src_2[0], 1);
+      BufferVectorType src_0_3 = svld1(pg, &src_3[0]);
+      BufferVectorType src_1_3 = svld1_vnum(pg, &src_3[0], 1);
+      BufferVectorType src_0_4 = svld1(pg, &src_4[0]);
+      BufferVectorType src_1_4 = svld1_vnum(pg, &src_4[0], 1);
+      BufferVectorType src_0_5 = svld1(pg, &src_5[0]);
+      BufferVectorType src_1_5 = svld1_vnum(pg, &src_5[0], 1);
+      BufferVectorType src_0_6 = svld1(pg, &src_6[0]);
+      BufferVectorType src_1_6 = svld1_vnum(pg, &src_6[0], 1);
+
+      filter_.horizontal_vector_path(pg, src_0_0, src_0_1, src_0_2, src_0_3,
+                                     src_0_4, src_0_5, src_0_6,
+                                     &dst_rows[index]);
+      filter_.horizontal_vector_path(
+          pg, src_1_0, src_1_1, src_1_2, src_1_3, src_1_4, src_1_5, src_1_6,
+          &dst_rows[index + BufferVecTraits::num_lanes()]);
+    } else if constexpr (KernelSize == 15) {
+      auto src_0 = &src_rows.at(0, border_offsets.c(0))[index];
+      auto src_1 = &src_rows.at(0, border_offsets.c(1))[index];
+      auto src_2 = &src_rows.at(0, border_offsets.c(2))[index];
+      auto src_3 = &src_rows.at(0, border_offsets.c(3))[index];
+      auto src_4 = &src_rows.at(0, border_offsets.c(4))[index];
+      auto src_5 = &src_rows.at(0, border_offsets.c(5))[index];
+      auto src_6 = &src_rows.at(0, border_offsets.c(6))[index];
+      auto src_7 = &src_rows.at(0, border_offsets.c(7))[index];
+      auto src_8 = &src_rows.at(0, border_offsets.c(8))[index];
+      auto src_9 = &src_rows.at(0, border_offsets.c(9))[index];
+      auto src_10 = &src_rows.at(0, border_offsets.c(10))[index];
+      auto src_11 = &src_rows.at(0, border_offsets.c(11))[index];
+      auto src_12 = &src_rows.at(0, border_offsets.c(12))[index];
+      auto src_13 = &src_rows.at(0, border_offsets.c(13))[index];
+      auto src_14 = &src_rows.at(0, border_offsets.c(14))[index];
+
+      BufferVectorType src_0_0 = svld1(pg, &src_0[0]);
+      BufferVectorType src_1_0 = svld1_vnum(pg, &src_0[0], 1);
+      BufferVectorType src_0_1 = svld1(pg, &src_1[0]);
+      BufferVectorType src_1_1 = svld1_vnum(pg, &src_1[0], 1);
+      BufferVectorType src_0_2 = svld1(pg, &src_2[0]);
+      BufferVectorType src_1_2 = svld1_vnum(pg, &src_2[0], 1);
+      BufferVectorType src_0_3 = svld1(pg, &src_3[0]);
+      BufferVectorType src_1_3 = svld1_vnum(pg, &src_3[0], 1);
+      BufferVectorType src_0_4 = svld1(pg, &src_4[0]);
+      BufferVectorType src_1_4 = svld1_vnum(pg, &src_4[0], 1);
+      BufferVectorType src_0_5 = svld1(pg, &src_5[0]);
+      BufferVectorType src_1_5 = svld1_vnum(pg, &src_5[0], 1);
+      BufferVectorType src_0_6 = svld1(pg, &src_6[0]);
+      BufferVectorType src_1_6 = svld1_vnum(pg, &src_6[0], 1);
+      BufferVectorType src_0_7 = svld1(pg, &src_7[0]);
+      BufferVectorType src_1_7 = svld1_vnum(pg, &src_7[0], 1);
+      BufferVectorType src_0_8 = svld1(pg, &src_8[0]);
+      BufferVectorType src_1_8 = svld1_vnum(pg, &src_8[0], 1);
+      BufferVectorType src_0_9 = svld1(pg, &src_9[0]);
+      BufferVectorType src_1_9 = svld1_vnum(pg, &src_9[0], 1);
+      BufferVectorType src_0_10 = svld1(pg, &src_10[0]);
+      BufferVectorType src_1_10 = svld1_vnum(pg, &src_10[0], 1);
+      BufferVectorType src_0_11 = svld1(pg, &src_11[0]);
+      BufferVectorType src_1_11 = svld1_vnum(pg, &src_11[0], 1);
+      BufferVectorType src_0_12 = svld1(pg, &src_12[0]);
+      BufferVectorType src_1_12 = svld1_vnum(pg, &src_12[0], 1);
+      BufferVectorType src_0_13 = svld1(pg, &src_13[0]);
+      BufferVectorType src_1_13 = svld1_vnum(pg, &src_13[0], 1);
+      BufferVectorType src_0_14 = svld1(pg, &src_14[0]);
+      BufferVectorType src_1_14 = svld1_vnum(pg, &src_14[0], 1);
+
+      filter_.horizontal_vector_path(
+          pg, src_0_0, src_0_1, src_0_2, src_0_3, src_0_4, src_0_5, src_0_6,
+          src_0_7, src_0_8, src_0_9, src_0_10, src_0_11, src_0_12, src_0_13,
+          src_0_14, &dst_rows[index]);
+      filter_.horizontal_vector_path(
+          pg, src_1_0, src_1_1, src_1_2, src_1_3, src_1_4, src_1_5, src_1_6,
+          src_1_7, src_1_8, src_1_9, src_1_10, src_1_11, src_1_12, src_1_13,
+          src_1_14, &dst_rows[index + BufferVecTraits::num_lanes()]);
+    } else {
+      static_assert(KernelSize != KernelSize,
+                    "please define variants for other kernel sizes");
+    }
   }
 
   template <size_t... SeqNum>
diff --git a/kleidicv/include/kleidicv/workspace/border.h b/kleidicv/include/kleidicv/workspace/border.h
index 3e5a8d34c0839c203de2da3e84d4861bb9f617b8..a1b4164bc20931a86eedfe0358bde3f906fa50f0 100644
--- a/kleidicv/include/kleidicv/workspace/border.h
+++ b/kleidicv/include/kleidicv/workspace/border.h
@@ -29,9 +29,15 @@ class FixedBorderInfo final {
       static_assert(sizeof...(args) == KernelSize);
     }
 
-    size_t c(int i) const { return offsets_[i]; }
+    size_t c(size_t i) const { return offsets_[i]; }
 
    private:
+    template <int... SeqNum>
+    static inline Offsets from_seq(std::integer_sequence<int, SeqNum...>) {
+      return Offsets{SeqNum...};
+    }
+
+    friend class FixedBorderInfo;
     size_t offsets_[KernelSize];
   };
 
@@ -50,19 +56,19 @@ class FixedBorderInfo final {
     constexpr auto seq = std::make_integer_sequence<int, (KernelSize >> 1)>{};
     switch (border_type_) {
       case FixedBorderType::REPLICATE:
-        return get_border<FixedBorderType::REPLICATE, false>(index, seq);
+        return get_offsets<FixedBorderType::REPLICATE, false>(index, seq);
         break;
 
       case FixedBorderType::REFLECT:
-        return get_border<FixedBorderType::REFLECT, false>(index, seq);
+        return get_offsets<FixedBorderType::REFLECT, false>(index, seq);
         break;
 
       case FixedBorderType::WRAP:
-        return get_border<FixedBorderType::WRAP, false>(index, seq);
+        return get_offsets<FixedBorderType::WRAP, false>(index, seq);
         break;
 
       case FixedBorderType::REVERSE:
-        return get_border<FixedBorderType::REVERSE, false>(index, seq);
+        return get_offsets<FixedBorderType::REVERSE, false>(index, seq);
         break;
     }
     // Unreachable. Compiler should emit a warning-as-error if any cases are
@@ -78,19 +84,19 @@ class FixedBorderInfo final {
     index = length_ - index - 1;
     switch (border_type_) {
       case FixedBorderType::REPLICATE:
-        return get_border<FixedBorderType::REPLICATE, true>(index, seq);
+        return get_offsets<FixedBorderType::REPLICATE, true>(index, seq);
         break;
 
       case FixedBorderType::REFLECT:
-        return get_border<FixedBorderType::REFLECT, true>(index, seq);
+        return get_offsets<FixedBorderType::REFLECT, true>(index, seq);
         break;
 
       case FixedBorderType::WRAP:
-        return get_border<FixedBorderType::WRAP, true>(index, seq);
+        return get_offsets<FixedBorderType::WRAP, true>(index, seq);
         break;
 
       case FixedBorderType::REVERSE:
-        return get_border<FixedBorderType::REVERSE, true>(index, seq);
+        return get_offsets<FixedBorderType::REVERSE, true>(index, seq);
         break;
     }
     // Unreachable. Compiler should emit a warning-as-error if any cases are
@@ -120,83 +126,135 @@ class FixedBorderInfo final {
       KLEIDICV_STREAMING_COMPATIBLE {
     // Example (15x15): Offsets{-7, -6, -5, -4, -3, -2, -1, 0, 1, 2, 3, 4, 5, 6,
     // 7};
-    return Offsets{(SeqNum - (KernelSize >> 1))..., 0, (SeqNum + 1)...};
+    constexpr int k_2 = static_cast<int>(KernelSize >> 1);
+    return Offsets{(SeqNum - k_2)..., 0, (SeqNum + 1)...};
   }
 
   // Creates the Offsets object containing offsets in various intervals
   // depending on the row/column index, border type as well the border
-  // position used. NOLINTBEGIN(readability-function-cognitive-complexity)
-  template <FixedBorderType BorderType, bool IsRight, int... SeqNum>
-  inline Offsets get_border(int index, std::integer_sequence<int, SeqNum...>)
-      const KLEIDICV_STREAMING_COMPATIBLE {
+  // position used.
+  template <int Index, FixedBorderType BorderType, bool IsRight, int... SeqNum>
+  static constexpr auto generate_border_offsets(
+      std::integer_sequence<int, SeqNum...>) {
+    constexpr int k_2 = static_cast<int>(KernelSize >> 1);
     if constexpr (BorderType == FixedBorderType::REPLICATE && !IsRight) {
       // Example (15x15, index 4, left): Offsets{-4, -4, -4, -4, -3, -2, -1, 0,
       // 1, 2, 3, 4, 5, 6, 7};
-      return Offsets{(SeqNum - static_cast<int>(KernelSize >> 1) < -index)
-                         ? -index
-                         : (SeqNum - (KernelSize >> 1))...,
-                     0, (SeqNum + 1)...};
+      return std::integer_sequence < int,
+             (SeqNum - k_2 < -Index) ? -Index : (SeqNum - k_2)..., 0,
+             (SeqNum + 1)... > {};
     }
 
     if constexpr (BorderType == FixedBorderType::REPLICATE && IsRight) {
       // Example (15x15, index 4, right): Offsets{-7, -6, -5, -4, -3, -2, -1, 0,
       // 1, 2, 3, 4, 4, 4, 4};
-      return Offsets{(SeqNum - (KernelSize >> 1))..., 0,
-                     (SeqNum >= index) ? index : (SeqNum + 1)...};
+      return std::integer_sequence < int, (SeqNum - k_2)..., 0,
+             (SeqNum >= Index) ? Index : (SeqNum + 1)... > {};
     }
 
     if constexpr (BorderType == FixedBorderType::REFLECT && !IsRight) {
       // Example (15x15, index 4, left): Offsets{-2, -3, -4, -4, -3, -2, -1, 0,
       // 1, 2, 3, 4, 5, 6, 7};
-      return Offsets{(SeqNum - static_cast<int>(KernelSize >> 1) < -index)
-                         ? ((KernelSize >> 1) - (index << 1) - (SeqNum + 1))
-                         : (SeqNum - (KernelSize >> 1))...,
-                     0, (SeqNum + 1)...};
+      return std::integer_sequence < int,
+             (SeqNum - k_2 < -Index) ? (k_2 - (Index << 1) - (SeqNum + 1))
+                                     : (SeqNum - k_2)...,
+             0, (SeqNum + 1)... > {};
     }
 
     if constexpr (BorderType == FixedBorderType::REFLECT && IsRight) {
       // Example (15x15, index 4, right): Offsets{-7, -6, -5, -4, -3, -2, -1, 0,
       // 1, 2, 3, 4, 4, 3, 2};
-      return Offsets{
-          (SeqNum - (KernelSize >> 1))..., 0,
-          (SeqNum >= index) ? ((index << 1) - SeqNum) : (SeqNum + 1)...};
+      return std::integer_sequence < int, (SeqNum - k_2)..., 0,
+             (SeqNum >= Index) ? ((Index << 1) - SeqNum) : (SeqNum + 1)... > {};
     }
 
-    if constexpr (BorderType == FixedBorderType::WRAP && !IsRight) {
-      // Example (15x15, index 4, left): Offsets{length_ - 7, length_ - 6,
-      // length_ - 5, -4, -3, -2, -1, 0, 1, 2, 3, 4, 5, 6, 7};
-      return Offsets{(SeqNum - static_cast<int>(KernelSize >> 1) < -index)
-                         ? (SeqNum - (KernelSize >> 1) + length_)
-                         : (SeqNum - (KernelSize >> 1))...,
-                     0, (SeqNum + 1)...};
+    if constexpr (BorderType == FixedBorderType::REVERSE && !IsRight) {
+      // Example (15x15, index 4, left): Offsets{-1, -2, -3, -4, -3, -2, -1, 0,
+      // 1, 2, 3, 4, 5, 6, 7};
+      return std::integer_sequence < int,
+             (SeqNum - k_2 < -Index) ? (k_2 - (Index << 1) - SeqNum)
+                                     : (SeqNum - k_2)...,
+             0, (SeqNum + 1)... > {};
     }
 
-    if constexpr (BorderType == FixedBorderType::WRAP && IsRight) {
+    if constexpr (BorderType == FixedBorderType::REVERSE && IsRight) {
       // Example (15x15, index 4, right): Offsets{-7, -6, -5, -4, -3, -2, -1, 0,
-      // 1, 2, 3, 4, 5 - length_, 6 - length_, 7 - length_};
-      return Offsets{
-          (SeqNum - (KernelSize >> 1))..., 0,
-          (SeqNum >= index) ? (SeqNum - length_ + 1) : (SeqNum + 1)...};
+      // 1, 2, 3, 4, 3, 2, 1};
+      return std::integer_sequence < int, (SeqNum - k_2)..., 0,
+             (SeqNum >= Index) ? ((Index << 1) - (SeqNum + 1))
+                               : (SeqNum + 1)... > {};
     }
+  }
 
-    if constexpr (BorderType == FixedBorderType::REVERSE && !IsRight) {
-      // Example (15x15, index 4, left): Offsets{-1, -2, -3, -4, -3, -2, -1, 0,
-      // 1, 2, 3, 4, 5, 6, 7};
-      return Offsets{(SeqNum - static_cast<int>(KernelSize >> 1) < -index)
-                         ? ((KernelSize >> 1) - (index << 1) - SeqNum)
-                         : (SeqNum - (KernelSize >> 1))...,
+  template <int Index, bool IsRight, int... SeqNum>
+  inline Offsets generate_border_offsets_wrap(
+      std::integer_sequence<int, SeqNum...>) const
+      KLEIDICV_STREAMING_COMPATIBLE {
+    constexpr int k_2 = static_cast<int>(KernelSize >> 1);
+    if constexpr (!IsRight) {
+      // Example (15x15, index 4, left): Offsets{length_ - 7, length_ - 6,
+      // length_ - 5, -4, -3, -2, -1, 0, 1, 2, 3, 4, 5, 6, 7};
+      return Offsets{(SeqNum - k_2 < -Index) ? (SeqNum - k_2 + length_)
+                                             : (SeqNum - k_2)...,
                      0, (SeqNum + 1)...};
     }
 
-    if constexpr (BorderType == FixedBorderType::REVERSE && IsRight) {
+    if constexpr (IsRight) {
       // Example (15x15, index 4, right): Offsets{-7, -6, -5, -4, -3, -2, -1, 0,
-      // 1, 2, 3, 4, 3, 2, 1};
+      // 1, 2, 3, 4, 5 - length_, 6 - length_, 7 - length_};
       return Offsets{
-          (SeqNum - (KernelSize >> 1))..., 0,
-          (SeqNum >= index) ? ((index << 1) - (SeqNum + 1)) : (SeqNum + 1)...};
+          (SeqNum - k_2)..., 0,
+          (SeqNum >= Index) ? (SeqNum - length_ + 1) : (SeqNum + 1)...};
+    }
+  }
+
+  // NOLINTBEGIN(readability-function-cognitive-complexity,
+  // readability-avoid-nested-conditional-operator)
+  template <FixedBorderType BorderType, bool IsRight, int... SeqNum>
+  inline Offsets get_offsets(int index,
+                             std::integer_sequence<int, SeqNum...> seq) const
+      KLEIDICV_STREAMING_COMPATIBLE {
+    if constexpr (BorderType == FixedBorderType::WRAP) {
+      return (index == 0)    ? generate_border_offsets_wrap<0, IsRight>(seq)
+             : (index == 1)  ? generate_border_offsets_wrap<1, IsRight>(seq)
+             : (index == 2)  ? generate_border_offsets_wrap<2, IsRight>(seq)
+             : (index == 3)  ? generate_border_offsets_wrap<3, IsRight>(seq)
+             : (index == 4)  ? generate_border_offsets_wrap<4, IsRight>(seq)
+             : (index == 5)  ? generate_border_offsets_wrap<5, IsRight>(seq)
+             : (index == 6)  ? generate_border_offsets_wrap<6, IsRight>(seq)
+             : (index == 7)  ? generate_border_offsets_wrap<7, IsRight>(seq)
+             : (index == 8)  ? generate_border_offsets_wrap<8, IsRight>(seq)
+             : (index == 9)  ? generate_border_offsets_wrap<9, IsRight>(seq)
+             : (index == 10) ? generate_border_offsets_wrap<10, IsRight>(seq)
+             : (index == 11) ? generate_border_offsets_wrap<11, IsRight>(seq)
+             : (index == 12) ? generate_border_offsets_wrap<12, IsRight>(seq)
+             : (index == 13) ? generate_border_offsets_wrap<13, IsRight>(seq)
+             : (index == 14) ? generate_border_offsets_wrap<14, IsRight>(seq)
+                             : Offsets{};
+    } else {
+      // clang-format off
+      return
+          (index == 0) ? Offsets::from_seq(generate_border_offsets<0, BorderType, IsRight>(seq))
+        : (index == 1) ? Offsets::from_seq(generate_border_offsets<1, BorderType, IsRight>(seq))
+        : (index == 2) ? Offsets::from_seq(generate_border_offsets<2, BorderType, IsRight>(seq))
+        : (index == 3) ? Offsets::from_seq(generate_border_offsets<3, BorderType, IsRight>(seq))
+        : (index == 4) ? Offsets::from_seq(generate_border_offsets<4, BorderType, IsRight>(seq))
+        : (index == 5) ? Offsets::from_seq(generate_border_offsets<5, BorderType, IsRight>(seq))
+        : (index == 6) ? Offsets::from_seq(generate_border_offsets<6, BorderType, IsRight>(seq))
+        : (index == 7) ? Offsets::from_seq(generate_border_offsets<7, BorderType, IsRight>(seq))
+        : (index == 8) ? Offsets::from_seq(generate_border_offsets<8, BorderType, IsRight>(seq))
+        : (index == 9) ? Offsets::from_seq(generate_border_offsets<9, BorderType, IsRight>(seq))
+        : (index == 10) ? Offsets::from_seq(generate_border_offsets<10, BorderType, IsRight>(seq))
+        : (index == 11) ? Offsets::from_seq(generate_border_offsets<11, BorderType, IsRight>(seq))
+        : (index == 12) ? Offsets::from_seq(generate_border_offsets<12, BorderType, IsRight>(seq))
+        : (index == 13) ? Offsets::from_seq(generate_border_offsets<13, BorderType, IsRight>(seq))
+        : (index == 14) ? Offsets::from_seq(generate_border_offsets<14, BorderType, IsRight>(seq))
+        : Offsets{};
+      // clang-format on
     }
   }
-  // NOLINTEND(readability-function-cognitive-complexity)
+  // NOLINTEND(readability-function-cognitive-complexity,
+  // readability-avoid-nested-conditional-operator)
 
   size_t length_;
   FixedBorderType border_type_;