diff --git a/CMakeLists.txt b/CMakeLists.txt
index 5fdf43ec62da32a922dce21af2d08fed7c2b2072..fe4f7fae41d02b753ac3edc35d8f8edea9d7df57 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -36,11 +36,11 @@ if(KLEIDIAI_ENABLE_CLANG_TIDY)
 endif()
 
 if(MSVC)
-    set(KLEIDIAI_WARNING_FLAGS_C
+    set(KLEIDIAI_WARNING_FLAGS_BASE
         # "/Wall" - This flag is disabled until the kernel library is cleaned up.
     )
 else()
-    set(KLEIDIAI_WARNING_FLAGS_C
+    set(KLEIDIAI_WARNING_FLAGS_BASE
         "-Wall"
         "-Wdisabled-optimization"
         "-Werror"
@@ -53,16 +53,24 @@ else()
         "-Wcast-qual"
     )
 
+    # C only flags not present in C++
+    set(KLEIDIAI_WARNING_FLAGS_C
+        "-Wmissing-prototypes"
+        "-Wstrict-prototypes"
+    )
+
     set(KLEIDIAI_WARNING_FLAGS_CXX
         "-Wctor-dtor-privacy"
         "-Weffc++"
         "-Woverloaded-virtual"
         "-Wsign-promo"
+        "-Wmissing-declarations"
     )
 endif()
 
 set(KLEIDIAI_WARNING_FLAGS
-    ${KLEIDIAI_WARNING_FLAGS_C}
+    ${KLEIDIAI_WARNING_FLAGS_BASE}
+    $<$<COMPILE_LANGUAGE:C>:${KLEIDIAI_WARNING_FLAGS_C}>
     $<$<COMPILE_LANGUAGE:CXX>:${KLEIDIAI_WARNING_FLAGS_CXX}>
 )
 
diff --git a/kai/ukernels/matmul/matmul_clamp_f16_f16p_f16p/kai_matmul_clamp_f16_f16p2vlx2_f16p2vlx2_2vlx2vl_sme2_mopa.c b/kai/ukernels/matmul/matmul_clamp_f16_f16p_f16p/kai_matmul_clamp_f16_f16p2vlx2_f16p2vlx2_2vlx2vl_sme2_mopa.c
index 6a130319715495ae00ac7391c385ec97dfccf1a1..a30936cf5dc2f98752b07617d5de2467ea588bb5 100644
--- a/kai/ukernels/matmul/matmul_clamp_f16_f16p_f16p/kai_matmul_clamp_f16_f16p2vlx2_f16p2vlx2_2vlx2vl_sme2_mopa.c
+++ b/kai/ukernels/matmul/matmul_clamp_f16_f16p_f16p/kai_matmul_clamp_f16_f16p2vlx2_f16p2vlx2_2vlx2vl_sme2_mopa.c
@@ -11,6 +11,8 @@
 #error This file must be compiled for AArch64, FEAT_SVE2.
 #else  // Architectural features check.
 
+#include "kai_matmul_clamp_f16_f16p2vlx2_f16p2vlx2_2vlx2vl_sme2_mopa.h"
+
 #include <arm_neon.h>
 #include <stddef.h>
 #include <stdint.h>
diff --git a/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa.c b/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa.c
index da28c6f0418ea8e8980fd4bd87e2a0d457eb8830..f677276fa943fadb800cbb8959ae29e583a7e4fb 100644
--- a/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa.c
+++ b/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa.c
@@ -11,6 +11,8 @@
 #error This file must be compiled for AArch64, FEAT_SVE2.
 #else  // Architectural features check.
 
+#include "kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa.h"
+
 #include <stddef.h>
 #include <stdint.h>
 
diff --git a/kai/ukernels/matmul/pack/kai_lhs_pack_bf16p2vlx2_f32_sme.c b/kai/ukernels/matmul/pack/kai_lhs_pack_bf16p2vlx2_f32_sme.c
index 351a4e3a1c393bf9621f3eee5f2f944003476989..e6992357e2723c441aa99d414ec0000b7694fc87 100644
--- a/kai/ukernels/matmul/pack/kai_lhs_pack_bf16p2vlx2_f32_sme.c
+++ b/kai/ukernels/matmul/pack/kai_lhs_pack_bf16p2vlx2_f32_sme.c
@@ -11,6 +11,8 @@
 #error This file must be compiled for AArch64, FEAT_SVE2.
 #else  // Architectural features check.
 
+#include "kai_lhs_pack_bf16p2vlx2_f32_sme.h"
+
 #include <stddef.h>
 #include <stdint.h>
 
diff --git a/kai/ukernels/matmul/pack/kai_lhs_pack_bf16p8x4_f16_neon.c b/kai/ukernels/matmul/pack/kai_lhs_pack_bf16p8x4_f16_neon.c
index 1cc0c0713a03fc921511632f8992c706f0df90f5..ce92ba1faf7c2969cc285147d2fdb026a36479e0 100644
--- a/kai/ukernels/matmul/pack/kai_lhs_pack_bf16p8x4_f16_neon.c
+++ b/kai/ukernels/matmul/pack/kai_lhs_pack_bf16p8x4_f16_neon.c
@@ -12,6 +12,8 @@
 #error This file must be compiled for AArch64, FEAT_BF16, FEAT_FP16.
 #else  // Architectural features check.
 
+#include "kai_lhs_pack_bf16p8x4_f16_neon.h"
+
 #include <stddef.h>
 #include <stdint.h>
 
diff --git a/kai/ukernels/matmul/pack/kai_lhs_pack_f32p2vlx1_f32_sme.c b/kai/ukernels/matmul/pack/kai_lhs_pack_f32p2vlx1_f32_sme.c
index d8b4708d8add870b8ad715c129b06d6029e0e2b3..e2636770a52284f43da864672edc0f2eb5ec10e5 100644
--- a/kai/ukernels/matmul/pack/kai_lhs_pack_f32p2vlx1_f32_sme.c
+++ b/kai/ukernels/matmul/pack/kai_lhs_pack_f32p2vlx1_f32_sme.c
@@ -11,6 +11,8 @@
 #error This file must be compiled for AArch64, FEAT_SVE2.
 #else  // Architectural features check.
 
+#include "kai_lhs_pack_f32p2vlx1_f32_sme.h"
+
 #include <stddef.h>
 #include <stdint.h>
 
diff --git a/kai/ukernels/matmul/pack/kai_lhs_pack_x16p2vlx2_x16_sme.c b/kai/ukernels/matmul/pack/kai_lhs_pack_x16p2vlx2_x16_sme.c
index f16d52b281ea96bf6333f86150951e554d76391a..d865117f190b0dc79aa14b12884634a3b83d20d7 100644
--- a/kai/ukernels/matmul/pack/kai_lhs_pack_x16p2vlx2_x16_sme.c
+++ b/kai/ukernels/matmul/pack/kai_lhs_pack_x16p2vlx2_x16_sme.c
@@ -11,6 +11,8 @@
 #error This file must be compiled for AArch64, FEAT_SVE2.
 #else  // Architectural features check.
 
+#include "kai_lhs_pack_x16p2vlx2_x16_sme.h"
+
 #include <stddef.h>
 #include <stdint.h>
 
diff --git a/kai/ukernels/matmul/pack/kai_lhs_quant_pack_bf16p1x4_f32_neon.c b/kai/ukernels/matmul/pack/kai_lhs_quant_pack_bf16p1x4_f32_neon.c
index 5d6f1587a107ec4dab40cc5941007227455228e9..fb0c4bd5d662d9030a61b6b3d3fbc1813e23fdf9 100644
--- a/kai/ukernels/matmul/pack/kai_lhs_quant_pack_bf16p1x4_f32_neon.c
+++ b/kai/ukernels/matmul/pack/kai_lhs_quant_pack_bf16p1x4_f32_neon.c
@@ -8,6 +8,8 @@
 #error This file must be compiled for AArch64, FEAT_BF16.
 #else  // Architectural features check.
 
+#include "kai_lhs_quant_pack_bf16p1x4_f32_neon.h"
+
 #include <arm_neon.h>
 #include <stddef.h>
 #include <stdint.h>
diff --git a/kai/ukernels/matmul/pack/kai_lhs_quant_pack_bf16p8x4_f32_neon.c b/kai/ukernels/matmul/pack/kai_lhs_quant_pack_bf16p8x4_f32_neon.c
index 9e59bdd0a79bc4b1794f49c0d21238797c89e433..60862dc4752105e42de65eedb9de55553000ccce 100644
--- a/kai/ukernels/matmul/pack/kai_lhs_quant_pack_bf16p8x4_f32_neon.c
+++ b/kai/ukernels/matmul/pack/kai_lhs_quant_pack_bf16p8x4_f32_neon.c
@@ -10,6 +10,8 @@
 
 #define MAX_MR 8
 
+#include "kai_lhs_quant_pack_bf16p8x4_f32_neon.h"
+
 #include <arm_neon.h>
 #include <stddef.h>
 #include <stdint.h>
@@ -50,8 +52,8 @@ size_t kai_get_lhs_packed_size_lhs_quant_pack_bf16p8x4_f32_neon(size_t m, size_t
 }
 
 void kai_run_lhs_quant_pack_bf16p8x4_f32_neon(
-    size_t m, size_t k, size_t mr, size_t kr, size_t sr, size_t m_idx_start, const float* lhs, size_t lhs_stride,
-    uint16_t* lhs_packed) {
+    size_t m, size_t k, size_t mr, size_t kr, size_t sr, size_t m_idx_start, const void* lhs, size_t lhs_stride,
+    void* lhs_packed) {
     KAI_ASSUME(mr == kai_mr);
     KAI_ASSUME(sr == kai_sr);
     KAI_ASSUME(kr == kai_kr);
diff --git a/kai/ukernels/matmul/pack/kai_rhs_pack_kxn_bf16p12x4biasf16_f16_neon.h b/kai/ukernels/matmul/pack/kai_rhs_pack_kxn_bf16p12x4biasf16_f16_neon.h
index cd3863021696c92ee0f37fb61c8dc39d263ea7d5..11a6601f302aba23d76d1f572d511edd5b8bfc52 100644
--- a/kai/ukernels/matmul/pack/kai_rhs_pack_kxn_bf16p12x4biasf16_f16_neon.h
+++ b/kai/ukernels/matmul/pack/kai_rhs_pack_kxn_bf16p12x4biasf16_f16_neon.h
@@ -1,5 +1,5 @@
 //
-// SPDX-FileCopyrightText: Copyright 2024 Arm Limited and/or its affiliates <open-source-office@arm.com>
+// SPDX-FileCopyrightText: Copyright 2024-2025 Arm Limited and/or its affiliates <open-source-office@arm.com>
 //
 // SPDX-License-Identifier: Apache-2.0
 //
@@ -33,18 +33,25 @@ size_t kai_get_rhs_offset_rhs_pack_kxn_bf16p12x4biasf16_f16_neon(size_t n_idx);
 /// @return The offset in bytes to the data element.
 size_t kai_get_bias_offset_rhs_pack_kxn_bf16p12x4biasf16_f16_neon(size_t n_idx);
 
+/// Gets the row stride in bytes to the packed RHS matrix.
+///
+/// @param[in] k Number of rows.
+///
+/// @return Row stride in bytes to the packed RHS matrix.
+size_t kai_get_rhs_packed_stride_rhs_pack_kxn_bf16p12x4biasf16_f16_neon(size_t k);
+
 /// Gets the offset in bytes to the data element in the packed RHS buffer.
 ///
-/// @param[in] n_idx Row index.
-/// @param[in] k Number of columns.
+/// @param[in] n_idx Column index.
+/// @param[in] k Number of rows.
 ///
 /// @return The offset in bytes to the data element.
 size_t kai_get_rhs_packed_offset_rhs_pack_kxn_bf16p12x4biasf16_f16_neon(size_t n_idx, size_t k);
 
 /// Gets the size in bytes of the packed RHS buffer.
 ///
-/// @param[in] n Number of rows.
-/// @param[in] k Number of columns.
+/// @param[in] n Number of columns.
+/// @param[in] k Number of rows.
 ///
 /// @return The size in bytes of the packed RHS buffer.
 size_t kai_get_rhs_packed_size_rhs_pack_kxn_bf16p12x4biasf16_f16_neon(size_t n, size_t k);
diff --git a/kai/ukernels/matmul/pack/kai_rhs_pack_kxn_bf16p12x4biasf32_f16_neon.c b/kai/ukernels/matmul/pack/kai_rhs_pack_kxn_bf16p12x4biasf32_f16_neon.c
index 1138a6790833e70c09f28fc05e94a1b9d6336165..5a431216126bc2e69639062198e984dfe48fc233 100644
--- a/kai/ukernels/matmul/pack/kai_rhs_pack_kxn_bf16p12x4biasf32_f16_neon.c
+++ b/kai/ukernels/matmul/pack/kai_rhs_pack_kxn_bf16p12x4biasf32_f16_neon.c
@@ -12,6 +12,8 @@
 #error This file must be compiled for AArch64, FEAT_BF16, FEAT_FP16.
 #else  // Architectural features check.
 
+#include "kai_rhs_pack_kxn_bf16p12x4biasf32_f16_neon.h"
+
 #include <arm_neon.h>
 #include <stddef.h>
 #include <stdint.h>
diff --git a/kai/ukernels/matmul/pack/kai_rhs_pack_kxn_f16p16x1biasf16_f16_f16_neon.c b/kai/ukernels/matmul/pack/kai_rhs_pack_kxn_f16p16x1biasf16_f16_f16_neon.c
index 61b8ba480ec595d630909b915f749a3ad3be6d2b..d85533d9ee920b4290e831acb560d5ddc3e7e3b7 100644
--- a/kai/ukernels/matmul/pack/kai_rhs_pack_kxn_f16p16x1biasf16_f16_f16_neon.c
+++ b/kai/ukernels/matmul/pack/kai_rhs_pack_kxn_f16p16x1biasf16_f16_f16_neon.c
@@ -1,5 +1,5 @@
 //
-// SPDX-FileCopyrightText: Copyright 2024 Arm Limited and/or its affiliates <open-source-office@arm.com>
+// SPDX-FileCopyrightText: Copyright 2024-2025 Arm Limited and/or its affiliates <open-source-office@arm.com>
 //
 // SPDX-License-Identifier: Apache-2.0
 //
@@ -8,6 +8,8 @@
 #error This file must be compiled for AArch64.
 #else  // Architectural features check.
 
+#include "kai_rhs_pack_kxn_f16p16x1biasf16_f16_f16_neon.h"
+
 #include <stddef.h>
 #include <stdint.h>
 
diff --git a/kai/ukernels/matmul/pack/kai_rhs_pack_kxn_f32p2vlx1biasf32_f32_f32_sme.c b/kai/ukernels/matmul/pack/kai_rhs_pack_kxn_f32p2vlx1biasf32_f32_f32_sme.c
index 5ce709fa256fba153637aaee3be76c991b64bb9f..55e4947460fb8493cd36b63f72c68c1c16f4755f 100644
--- a/kai/ukernels/matmul/pack/kai_rhs_pack_kxn_f32p2vlx1biasf32_f32_f32_sme.c
+++ b/kai/ukernels/matmul/pack/kai_rhs_pack_kxn_f32p2vlx1biasf32_f32_f32_sme.c
@@ -1,5 +1,5 @@
 //
-// SPDX-FileCopyrightText: Copyright 2024 Arm Limited and/or its affiliates <open-source-office@arm.com>
+// SPDX-FileCopyrightText: Copyright 2024-2025 Arm Limited and/or its affiliates <open-source-office@arm.com>
 //
 // SPDX-License-Identifier: Apache-2.0
 //
@@ -8,6 +8,8 @@
 #error This file must be compiled for AArch64, FEAT_SVE2.
 #else  // Architectural features check.
 
+#include "kai_rhs_pack_kxn_f32p2vlx1biasf32_f32_f32_sme.h"
+
 #include <stddef.h>
 #include <stdint.h>
 
diff --git a/kai/ukernels/matmul/pack/kai_rhs_pack_kxn_f32p8x1biasf32_f32_f32_neon.c b/kai/ukernels/matmul/pack/kai_rhs_pack_kxn_f32p8x1biasf32_f32_f32_neon.c
index 0c6b0074b2c751c414112861a2864d7a34604dee..afa3d8b597cd510e9116933a14c26da1b62ddbb6 100644
--- a/kai/ukernels/matmul/pack/kai_rhs_pack_kxn_f32p8x1biasf32_f32_f32_neon.c
+++ b/kai/ukernels/matmul/pack/kai_rhs_pack_kxn_f32p8x1biasf32_f32_f32_neon.c
@@ -1,5 +1,5 @@
 //
-// SPDX-FileCopyrightText: Copyright 2024 Arm Limited and/or its affiliates <open-source-office@arm.com>
+// SPDX-FileCopyrightText: Copyright 2024-2025 Arm Limited and/or its affiliates <open-source-office@arm.com>
 //
 // SPDX-License-Identifier: Apache-2.0
 //
@@ -8,6 +8,8 @@
 #error This file must be compiled for AArch64.
 #else  // Architectural features check.
 
+#include "kai_rhs_pack_kxn_f32p8x1biasf32_f32_f32_neon.h"
+
 #include <stddef.h>
 #include <stdint.h>
 
diff --git a/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_f32p2vlx1biasf32_f32_f32_sme.c b/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_f32p2vlx1biasf32_f32_f32_sme.c
index 426bfc6cbc9dc836c8d05082ea60180df415634d..b5fe9904812c5078bc799ec16bf56efdda568fbe 100644
--- a/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_f32p2vlx1biasf32_f32_f32_sme.c
+++ b/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_f32p2vlx1biasf32_f32_f32_sme.c
@@ -11,6 +11,8 @@
 #error This file must be compiled for AArch64, FEAT_SVE2.
 #else  // Architectural features check.
 
+#include "kai_rhs_pack_nxk_f32p2vlx1biasf32_f32_f32_sme.h"
+
 #include <stddef.h>
 #include <stdint.h>
 
diff --git a/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_f32p2vlx1biasf32_f32_f32_sme.h b/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_f32p2vlx1biasf32_f32_f32_sme.h
index 8bd3f6ee783104b625724cf524b9e471145ecbe2..77e68f6987ccd0aee9841834fc9ed4628de00c93 100644
--- a/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_f32p2vlx1biasf32_f32_f32_sme.h
+++ b/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_f32p2vlx1biasf32_f32_f32_sme.h
@@ -1,5 +1,5 @@
 //
-// SPDX-FileCopyrightText: Copyright 2024 Arm Limited and/or its affiliates <open-source-office@arm.com>
+// SPDX-FileCopyrightText: Copyright 2024-2025 Arm Limited and/or its affiliates <open-source-office@arm.com>
 //
 // SPDX-License-Identifier: Apache-2.0
 //
@@ -22,10 +22,10 @@ size_t kai_get_n_step_rhs_pack_nxk_f32p2vlx1biasf32_f32_f32_sme(void);
 /// Gets the offset in bytes to the data element in the RHS matrix buffer.
 ///
 /// @param[in] n_idx Column index.
-/// @param[in] rhs_offset Row stride in bytes of the RHS matrix.
+/// @param[in] rhs_stride Row stride in bytes of the RHS matrix.
 ///
 /// @return The offset in bytes to the data element.
-size_t kai_get_rhs_offset_rhs_pack_nxk_f32p2vlx1biasf32_f32_f32_sme(size_t n_idx, size_t rhs_offset);
+size_t kai_get_rhs_offset_rhs_pack_nxk_f32p2vlx1biasf32_f32_f32_sme(size_t n_idx, size_t rhs_stride);
 
 /// Gets the offset in bytes to the data element in the bias buffer.
 ///
@@ -34,6 +34,13 @@ size_t kai_get_rhs_offset_rhs_pack_nxk_f32p2vlx1biasf32_f32_f32_sme(size_t n_idx
 /// @return The offset in bytes to the data element.
 size_t kai_get_bias_offset_rhs_pack_nxk_f32p2vlx1biasf32_f32_f32_sme(size_t n_idx);
 
+/// Gets the row stride in bytes to the packed RHS matrix.
+///
+/// @param[in] k Number of columns.
+///
+/// @return Row stride in bytes to the packed RHS matrix.
+size_t kai_get_rhs_packed_stride_rhs_pack_nxk_f32p2vlx1biasf32_f32_f32_sme(size_t k);
+
 /// Gets the offset in bytes to the data element in the packed RHS buffer.
 ///
 /// @param[in] n_idx Row index.
diff --git a/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.h b/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.h
index f06e736cb61ed64cdc6858127cc169111ad09b8b..a350c0af6cf7837a3a41866b008ba1475ecdcab0 100644
--- a/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.h
+++ b/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.h
@@ -1,5 +1,5 @@
 //
-// SPDX-FileCopyrightText: Copyright 2024 Arm Limited and/or its affiliates <open-source-office@arm.com>
+// SPDX-FileCopyrightText: Copyright 2024-2025 Arm Limited and/or its affiliates <open-source-office@arm.com>
 //
 // SPDX-License-Identifier: Apache-2.0
 //
@@ -14,6 +14,17 @@
 extern "C" {
 #endif
 
+/// Gets the row stride in bytes to the packed RHS matrix.
+///
+/// @param[in] k The number of columns in the RHS matrix (not packed).
+/// @param[in] nr The number of columns written by the matmul micro-kernel.
+/// @param[in] kr The number of columns loaded in the single innermost loop of the matmul micro-kernel.
+/// @param[in] bl The block length, which defines the number of K values stored in a single block. It must be equivalent
+/// to 32.
+///
+/// @return Row stride in bytes to the packed RHS matrix.
+size_t kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0(size_t k, size_t nr, size_t kr, size_t bl);
+
 /// Gets the offset in bytes for the RHS matrix (not packed), which holds
 /// the int4 values in a N x K matrix, where N is number of rows and K is the number of columns.
 ///
diff --git a/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_x16p2vlx2b_x16_x16_sme.c b/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_x16p2vlx2b_x16_x16_sme.c
index 8ca803bd188a082f1d324c1f0ccc99f33d20cd4b..d292185823789d084ec41733df292b8260bec413 100644
--- a/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_x16p2vlx2b_x16_x16_sme.c
+++ b/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_x16p2vlx2b_x16_x16_sme.c
@@ -11,6 +11,8 @@
 #error This file must be compiled for AArch64, FEAT_SVE2.
 #else  // Architectural features check.
 
+#include "kai_rhs_pack_nxk_x16p2vlx2b_x16_x16_sme.h"
+
 #include <stddef.h>
 #include <stdint.h>
 
diff --git a/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_x16p2vlx2b_x16_x16_sme.h b/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_x16p2vlx2b_x16_x16_sme.h
index e5cbc7915ca1e83424625bfaedb1f9af1d315cc6..f9ad4b7f5b9dc97e591ccdc0542206e9bee2a323 100644
--- a/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_x16p2vlx2b_x16_x16_sme.h
+++ b/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_x16p2vlx2b_x16_x16_sme.h
@@ -1,5 +1,5 @@
 //
-// SPDX-FileCopyrightText: Copyright 2024 Arm Limited and/or its affiliates <open-source-office@arm.com>
+// SPDX-FileCopyrightText: Copyright 2024-2025 Arm Limited and/or its affiliates <open-source-office@arm.com>
 //
 // SPDX-License-Identifier: Apache-2.0
 //
@@ -22,10 +22,10 @@ size_t kai_get_n_step_rhs_pack_nxk_x16p2vlx2b_x16_x16_sme(void);
 /// Gets the offset in bytes to the data element in the RHS matrix buffer.
 ///
 /// @param[in] n_idx Column index.
-/// @param[in] rhs_offset Row stride in bytes of the RHS matrix.
+/// @param[in] rhs_stride Row stride in bytes of the RHS matrix.
 ///
 /// @return The offset in bytes to the data element.
-size_t kai_get_rhs_offset_rhs_pack_nxk_x16p2vlx2b_x16_x16_sme(size_t n_idx, size_t rhs_offset);
+size_t kai_get_rhs_offset_rhs_pack_nxk_x16p2vlx2b_x16_x16_sme(size_t n_idx, size_t rhs_stride);
 
 /// Gets the offset in bytes to the data element in the bias buffer.
 ///
diff --git a/kai/ukernels/matmul/pack/kai_rhs_quant_pack_kxn_bf16p12x4biasf32_f32_neon.c b/kai/ukernels/matmul/pack/kai_rhs_quant_pack_kxn_bf16p12x4biasf32_f32_neon.c
index 78b7f9cd9803452b90e2bfaa5993ee49edf44333..e08021d0eae140d08b33956cc809636b8d84cd4c 100644
--- a/kai/ukernels/matmul/pack/kai_rhs_quant_pack_kxn_bf16p12x4biasf32_f32_neon.c
+++ b/kai/ukernels/matmul/pack/kai_rhs_quant_pack_kxn_bf16p12x4biasf32_f32_neon.c
@@ -13,6 +13,8 @@
 
 #define MAX_NR 12
 
+#include "kai_rhs_quant_pack_kxn_bf16p12x4biasf32_f32_neon.h"
+
 #include <arm_neon.h>
 #include <stddef.h>
 #include <stdint.h>
@@ -24,9 +26,8 @@ static const size_t kai_nr = 12;
 static const size_t kai_kr = 4;
 static const size_t kai_sr = 1;
 
-size_t kai_get_n_step_rhs_quant_pack_kxn_bf16p12x4biasf32_f32_neon(size_t nr) {
-    KAI_ASSUME(kai_nr == nr);
-    return nr;
+size_t kai_get_n_step_rhs_quant_pack_kxn_bf16p12x4biasf32_f32_neon(void) {
+    return kai_nr;
 }
 
 size_t kai_get_rhs_offset_rhs_quant_pack_kxn_bf16p12x4biasf32_f32_neon(size_t n_idx) {
diff --git a/kai_defs.bzl b/kai_defs.bzl
index 66ec4414ca4e9e4501936d171d0f84b63fc94e4a..3848ba2b762a3860a3ca8e58224193f045ec5ce3 100644
--- a/kai_defs.bzl
+++ b/kai_defs.bzl
@@ -28,12 +28,20 @@ def kai_gcc_warn_copts():
         "-Wcast-qual",
     ]
 
+# GCC/CLANG C only warning options
+def kai_gcc_warn_conlyopts():
+    return [
+        "-Wmissing-prototypes",
+        "-Wstrict-prototypes",
+    ]
+
 def kai_gcc_warn_cxxopts():
     return kai_gcc_warn_copts() + [
         "-Wctor-dtor-privacy",
         "-Weffc++",
         "-Woverloaded-virtual",
         "-Wsign-promo",
+        "-Wmissing-declarations",
     ]
 
 # GCC/CLANG compiler options
@@ -142,6 +150,7 @@ def _kai_c_cxx_common(name, copts_def_func, **kwargs):
         extra_copts.append("-fno-tree-vectorize")
 
     kwargs["copts"] = kwargs.get("copts", []) + copts_def_func(cpu_uarch) + extra_copts
+    kwargs["conlyopts"] = kai_gcc_warn_conlyopts()
     kwargs["deps"] = ["//:common"] + kwargs.get("deps", [])
     kwargs["linkstatic"] = kwargs.get("linkstatic", True)
 
diff --git a/test/common/printer.cpp b/test/common/printer.cpp
index 9fc3e0d8fb096f414b2df8d87e36efba6c23e940..eaad08a470b98d5acd10117fee72eff5402045c1 100644
--- a/test/common/printer.cpp
+++ b/test/common/printer.cpp
@@ -1,9 +1,11 @@
 //
-// SPDX-FileCopyrightText: Copyright 2024 Arm Limited and/or its affiliates <open-source-office@arm.com>
+// SPDX-FileCopyrightText: Copyright 2024-2025 Arm Limited and/or its affiliates <open-source-office@arm.com>
 //
 // SPDX-License-Identifier: Apache-2.0
 //
 
+#include "test/common/printer.hpp"
+
 #include <cstddef>
 #include <cstdint>
 #include <ostream>
diff --git a/test/reference/cast.cpp b/test/reference/cast.cpp
index 487c2f1e0f4e7b897e4a809d1c941707d3676281..67ad0e779650e38bcc160df5dfcacd37d329a211 100644
--- a/test/reference/cast.cpp
+++ b/test/reference/cast.cpp
@@ -1,9 +1,11 @@
 //
-// SPDX-FileCopyrightText: Copyright 2024 Arm Limited and/or its affiliates <open-source-office@arm.com>
+// SPDX-FileCopyrightText: Copyright 2024-2025 Arm Limited and/or its affiliates <open-source-office@arm.com>
 //
 // SPDX-License-Identifier: Apache-2.0
 //
 
+#include "test/reference/cast.hpp"
+
 #include <cstddef>
 #include <cstdint>
 #include <vector>
diff --git a/test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp b/test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp
index 638b1dc3364ae880e45f9cf8ad21b4115c0a73f6..81ffaf4d555e7c5396f7409c313ed6e0b931d4a4 100644
--- a/test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp
+++ b/test/tests/matmul_clamp_qai8_qai8p_qsi8cxp_test.cpp
@@ -1,5 +1,5 @@
 //
-// SPDX-FileCopyrightText: Copyright 2024 Arm Limited and/or its affiliates <open-source-office@arm.com>
+// SPDX-FileCopyrightText: Copyright 2024-2025 Arm Limited and/or its affiliates <open-source-office@arm.com>
 //
 // SPDX-License-Identifier: Apache-2.0
 //
@@ -182,7 +182,7 @@ struct TestReference {
 };
 
 /// Generate test reference data
-TestReference get_test_reference(const MatMulShape& shape, const MatMulVariant& variant) {
+static TestReference get_test_reference(const MatMulShape& shape, const MatMulVariant& variant) {
     // ============================================================
     // Generates input and reference output data
     // ============================================================
@@ -282,7 +282,7 @@ TestReference get_test_reference(const MatMulShape& shape, const MatMulVariant&
 }
 
 /// Test LHS packing
-void test_lhs_pack(
+static void test_lhs_pack(
     const MatMulShape& shape, const MatMulVariant& variant, const Rect& output_area, const TestReference& reference) {
     KAI_ASSUME(variant.lhs_pack.has_value());
 
@@ -313,7 +313,7 @@ void test_lhs_pack(
 }
 
 /// Test RHS packing
-void test_rhs_pack(
+static void test_rhs_pack(
     const MatMulShape& shape, const MatMulVariant& variant, const Rect& output_area, const TestReference& reference) {
     const auto imp_packed_rhs_size = variant.rhs_pack.get_packed_rhs_size(shape.n, shape.k);
     ASSERT_EQ(imp_packed_rhs_size, reference.packed_rhs.size());
@@ -349,7 +349,7 @@ void test_rhs_pack(
 }
 
 /// Test MatMul of GEMM like kernel
-void test_matmul(
+static void test_matmul(
     const MatMulShape& shape, const MatMulVariant& variant, const Rect& output_area, const TestReference& reference) {
     const auto imp_dst_size = variant.matmul.get_dst_size(shape.m, shape.n);
     ASSERT_EQ(imp_dst_size, reference.dst_qsi8_clamped.size());