void test_vst2Qu8 (void)
{
uint8_t *arg0_uint8_t;
uint8x16x2_t arg1_uint8x16x2_t;
vst2q_u8 (arg0_uint8_t, arg1_uint8x16x2_t);
}
/* Routine optimized for unshuffling a buffer for a type size of 2 bytes. */
static void
unshuffle2_neon(uint8_t * const
dest,
const uint8_t* const src,
const size_t vectorizable_elements,
const size_t total_elements
)
{
size_t i, j, k;
static const size_t bytesoftype = 2;
uint8x16x2_t r0;
for(
i = 0, k = 0;
i<vectorizable_elements*bytesoftype;
i += 32, k++) {
/* Load 32 bytes to the structure r0 */
for(
j = 0;
j < 2; j++) {
r0.val[j] =
vld1q_u8(src
+
total_elements* j
+ k*16);
}
/* Store (with permutation) the results in the destination vector */
vst2q_u8(dest
+ k*32, r0);
}
}
inline void vst2q(u8 * ptr, const uint8x16x2_t & v) { return vst2q_u8(ptr, v); }
