memcpy 16-bit values to 32-bit values (zero extend using SSE4.1)

memcpy_16to32.c

#include <smmintrin.h>

void memcpy_16to32(uint32_t *pdst_, const uint16_t *psrc_, size_t n)
{
	size_t items = n / sizeof(uint16_t);
  
	__m128i *psrc = (__m128i *) psrc_;
	__m128i *pdst = (__m128i *) pdst_;

	while (items >= 8) {
		// Read 8x 16-bit
		__m128i src = *psrc++;
		// Zero extend 4 first elements from 16-bit to 32-bit
		*pdst++ = _mm_cvtepu16_epi32(src);
		// Shift right 8 bytes
		src = _mm_srli_si128(src, 8);
		// Zero extend 4 last elements from 16-bit to 32-bit
		*pdst++ = _mm_cvtepu16_epi32(src);
    
		items -= 8;
	}

	// Copy last 7 or less items using for-loop
	uint16_t *psrc_u16 = (uint16_t*) psrc;
	uint32_t *pdst_u32 = (uint32_t*) pdst;
	while (items-- > 0) {
		*pdst_u32++ = *psrc_u16++;
	}
}