Fix NEON vs C mismatch of aom_hadamard_32x32() function

This CL resolves the mismatch between C and NEON implementation
of aom_hadamard_32x32 function.

Bug: aomedia:3455

Change-Id: I2f05890bd4e2a323f82810484720f982704dd3b6
(cherry picked from commit e752733bd9eaf700caf27765b107c8fa8439336e)

aom_dsp/arm/hadamard_neon.c

diff --git a/aom_dsp/arm/hadamard_neon.c b/aom_dsp/arm/hadamard_neon.c
index eda5db0..93f2b81 100644
--- a/aom_dsp/arm/hadamard_neon.c
+++ b/aom_dsp/arm/hadamard_neon.c
@@ -271,15 +271,15 @@
     const int16x8_t a2 = load_tran_low_to_s16q(coeff + 512);
     const int16x8_t a3 = load_tran_low_to_s16q(coeff + 768);
 
-    const int16x8_t b0 = vshrq_n_s16(vaddq_s16(a0, a1), 2);
-    const int16x8_t b1 = vshrq_n_s16(vsubq_s16(a0, a1), 2);
-    const int16x8_t b2 = vshrq_n_s16(vaddq_s16(a2, a3), 2);
-    const int16x8_t b3 = vshrq_n_s16(vsubq_s16(a2, a3), 2);
+    const int16x8_t b0 = vhaddq_s16(a0, a1);
+    const int16x8_t b1 = vhsubq_s16(a0, a1);
+    const int16x8_t b2 = vhaddq_s16(a2, a3);
+    const int16x8_t b3 = vhsubq_s16(a2, a3);
 
-    const int16x8_t c0 = vaddq_s16(b0, b2);
-    const int16x8_t c1 = vaddq_s16(b1, b3);
-    const int16x8_t c2 = vsubq_s16(b0, b2);
-    const int16x8_t c3 = vsubq_s16(b1, b3);
+    const int16x8_t c0 = vhaddq_s16(b0, b2);
+    const int16x8_t c1 = vhaddq_s16(b1, b3);
+    const int16x8_t c2 = vhsubq_s16(b0, b2);
+    const int16x8_t c3 = vhsubq_s16(b1, b3);
 
     store_s16q_to_tran_low(coeff + 0, c0);
     store_s16q_to_tran_low(coeff + 256, c1);