+void memcpy_interleaved_word_slow(uint16_t *dest1, uint16_t *dest2, const uint16_t *src, size_t n)
+{
+ assert(n % 2 == 0);
+ uint16_t *dptr1 = dest1;
+ uint16_t *dptr2 = dest2;
+
+ for (size_t i = 0; i < n; i += 2) {
+ *dptr1++ = *src++;
+ *dptr2++ = *src++;
+ }
+}
+
+#if HAS_MULTIVERSIONING
+
+// uint8_t version.
+
+__attribute__((target("default")))
+size_t memcpy_interleaved_fastpath_core(uint8_t *dest1, uint8_t *dest2, const uint8_t *src, const uint8_t *limit);
+
+__attribute__((target("avx2")))
+size_t memcpy_interleaved_fastpath_core(uint8_t *dest1, uint8_t *dest2, const uint8_t *src, const uint8_t *limit);
+
+__attribute__((target("default")))
+size_t memcpy_interleaved_fastpath_core(uint8_t *dest1, uint8_t *dest2, const uint8_t *src, const uint8_t *limit)
+{
+ // No fast path supported unless we have AVX2.
+ return 0;
+}
+
+__attribute__((target("avx2")))
+size_t memcpy_interleaved_fastpath_core(uint8_t *dest1, uint8_t *dest2, const uint8_t *src, const uint8_t *limit)
+{
+ size_t consumed = 0;
+ const __m256i *__restrict in = (const __m256i *)src;
+ __m256i *__restrict out1 = (__m256i *)dest1;
+ __m256i *__restrict out2 = (__m256i *)dest2;
+
+ __m256i shuffle_cw = _mm256_set_epi8(
+ 15, 13, 11, 9, 7, 5, 3, 1, 14, 12, 10, 8, 6, 4, 2, 0,
+ 15, 13, 11, 9, 7, 5, 3, 1, 14, 12, 10, 8, 6, 4, 2, 0);
+ while (in < (const __m256i *)limit) {
+ // Note: For brevity, comments show lanes as if they were 2x64-bit (they're actually 2x128).
+ __m256i data1 = _mm256_stream_load_si256(in); // AaBbCcDd EeFfGgHh
+ __m256i data2 = _mm256_stream_load_si256(in + 1); // IiJjKkLl MmNnOoPp
+
+ data1 = _mm256_shuffle_epi8(data1, shuffle_cw); // ABCDabcd EFGHefgh
+ data2 = _mm256_shuffle_epi8(data2, shuffle_cw); // IJKLijkl MNOPmnop
+
+ data1 = _mm256_permute4x64_epi64(data1, 0b11011000); // ABCDEFGH abcdefgh
+ data2 = _mm256_permute4x64_epi64(data2, 0b11011000); // IJKLMNOP ijklmnop
+
+ __m256i lo = _mm256_permute2x128_si256(data1, data2, 0b00100000);
+ __m256i hi = _mm256_permute2x128_si256(data1, data2, 0b00110001);
+
+ _mm256_storeu_si256(out1, lo);
+ _mm256_storeu_si256(out2, hi);
+
+ in += 2;
+ ++out1;
+ ++out2;
+ consumed += 64;
+ }
+
+ return consumed;
+}