From 61389b5e510f634943b2c68713cdae0ca67665e2 Mon Sep 17 00:00:00 2001
From: Anna Khakimova <anna.khakimova@intel.com>
Date: Thu, 11 Mar 2021 12:13:11 +0300
Subject: [PATCH] Fix for compilation issue for ARM 32-bit (#4717)

---
 .../ie_preprocess_gapi_kernels_neon.cpp       |  8 ++---
 .../thirdparty/ocv/opencv_hal_neon.hpp        | 35 ++++++++-----------
 2 files changed, 19 insertions(+), 24 deletions(-)

diff --git a/inference-engine/src/preprocessing/arm_neon/ie_preprocess_gapi_kernels_neon.cpp b/inference-engine/src/preprocessing/arm_neon/ie_preprocess_gapi_kernels_neon.cpp
index 0ee824c8748..b4dcb20213b 100644
--- a/inference-engine/src/preprocessing/arm_neon/ie_preprocess_gapi_kernels_neon.cpp
+++ b/inference-engine/src/preprocessing/arm_neon/ie_preprocess_gapi_kernels_neon.cpp
@@ -257,8 +257,8 @@ CV_ALWAYS_INLINE void horizontal_4LPI(std::array<std::array<uint8_t*, 4>, chanNu
                 v_uint8 q2 = v_shuffle(q0, hmask);
                 v_uint8 q3 = v_shuffle(q1, hmask);
 
-                v_uint8 q4 = v_blend<0xCC /*0b11001100*/>(q2, v_slli_si128(q3, 4));
-                v_uint8 q5 = v_blend<0xCC /*0b11001100*/>(v_srli_si128(q2, 4), q3);
+                v_uint8 q4 = v_blend<0xCC /*0b11001100*/>(q2, v_shift_left<4>(q3));
+                v_uint8 q5 = v_blend<0xCC /*0b11001100*/>(v_shift_right<4>(q2), q3);
 
                 v_store_low(&dst[c][0][x], q4);
                 v_store_high(&dst[c][1][x], q4);
@@ -334,8 +334,8 @@ CV_ALWAYS_INLINE void calcRowLinear_8UC_Impl_(std::array<std::array<uint8_t*, 4>
                     v_uint8 q0 = v_pack_u(r0, r1);
                     v_uint8 q1 = v_pack_u(r2, r3);
 
-                    v_uint8 q2 = v_blend<0xCC /*0b11001100*/>(q0, v_slli_si128(q1, 4));
-                    v_uint8 q3 = v_blend<0xCC /*0b11001100*/>(v_srli_si128(q0, 4), q1);
+                    v_uint8 q2 = v_blend<0xCC /*0b11001100*/>(q0, v_shift_left<4>(q1));
+                    v_uint8 q3 = v_blend<0xCC /*0b11001100*/>(v_shift_right<4>(q0), q1);
 
                     v_uint8 q4 = v_shuffle(q2, vmask);
                     v_uint8 q5 = v_shuffle(q3, vmask);
diff --git a/inference-engine/thirdparty/ocv/opencv_hal_neon.hpp b/inference-engine/thirdparty/ocv/opencv_hal_neon.hpp
index d37d377b012..0427423d788 100644
--- a/inference-engine/thirdparty/ocv/opencv_hal_neon.hpp
+++ b/inference-engine/thirdparty/ocv/opencv_hal_neon.hpp
@@ -46,6 +46,7 @@
 #define OPENCV_HAL_INTRIN_NEON_HPP
 
 #include <algorithm>
+#include <cassert>
 
 
 namespace cv
@@ -2447,34 +2448,28 @@ CV_ALWAYS_INLINE v_uint8x16 v_shuffle(const v_uint8x16& a, const v_uint8x16& mas
 #endif
 }
 
-CV_ALWAYS_INLINE v_uint8x16 v_slli_si128(const v_uint8x16& a, const int imm)
+template<int shift>
+CV_ALWAYS_INLINE v_uint8x16 v_slli_si128(const v_uint8x16& a)
 {
-    uint8x16_t ret = {};
-    if (imm <= 0) {
-        ret = a.val;
-    }
-    if (imm > 15) {
-        ret = vdupq_n_u8(0);
-    } else {
-        ret = vextq_u8(vdupq_n_u8(0), a.val, 16 - (imm));
-    }
+    assert((shift > 0) && (shift <= 15));
+    uint8x16_t ret = vextq_u8(vdupq_n_u8(0), a.val, shift /*16 - (imm)*/);
     return v_uint8x16(ret);
 }
 
-CV_ALWAYS_INLINE v_uint8x16 v_srli_si128(const v_uint8x16& a, const int imm)
+template<int shift>
+CV_ALWAYS_INLINE v_uint8x16 v_shift_right(const v_uint8x16& a)
 {
-    uint8x16_t ret = {};
-    if (imm <= 0) {
-        ret = a.val;
-    }
-    if (imm > 15) {
-        ret = vdupq_n_u8(0);
-    } else {
-        ret = vextq_u8(a.val, vdupq_n_u8(0), imm);
-    }
+    assert((shift > 0) && (shift <= 15));
+    uint8x16_t ret = vextq_u8(a.val, vdupq_n_u8(0), shift);
     return v_uint8x16(ret);
 }
 
+template<int shift>
+CV_ALWAYS_INLINE v_uint8x16 v_shift_left(const v_uint8x16& a)
+{
+    return v_slli_si128<16 - shift>(a);
+}
+
 CV_CPU_OPTIMIZATION_HAL_NAMESPACE_END
 
 //! @endcond