static void hor_add_sub_avx2(__m128i *row0, __m128i *row1){
__m128i a = _mm_hadd_epi16(*row0, *row1);
__m128i b = _mm_hsub_epi16(*row0, *row1);
__m128i c = _mm_hadd_epi16(a, b);
__m128i d = _mm_hsub_epi16(a, b);
*row0 = _mm_hadd_epi16(c, d);
*row1 = _mm_hsub_epi16(c, d);
}
/* Test the 128-bit form */
static void
ssse3_test_phsubw128 (int *i1, int *i2, int *r)
{
/* Assumes incoming pointers are 16-byte aligned */
__m128i t1 = *(__m128i *) i1;
__m128i t2 = *(__m128i *) i2;
*(__m128i *) r = _mm_hsub_epi16 (t1, t2);
}
static unsigned satd_8bit_4x4_avx2(const kvz_pixel *org, const kvz_pixel *cur)
{
__m128i original = _mm_cvtepu8_epi16(_mm_loadl_epi64((__m128i*)org));
__m128i current = _mm_cvtepu8_epi16(_mm_loadl_epi64((__m128i*)cur));
__m128i diff_lo = _mm_sub_epi16(current, original);
original = _mm_cvtepu8_epi16(_mm_loadl_epi64((__m128i*)(org + 8)));
current = _mm_cvtepu8_epi16(_mm_loadl_epi64((__m128i*)(cur + 8)));
__m128i diff_hi = _mm_sub_epi16(current, original);
//Hor
__m128i row0 = _mm_hadd_epi16(diff_lo, diff_hi);
__m128i row1 = _mm_hsub_epi16(diff_lo, diff_hi);
__m128i row2 = _mm_hadd_epi16(row0, row1);
__m128i row3 = _mm_hsub_epi16(row0, row1);
//Ver
row0 = _mm_hadd_epi16(row2, row3);
row1 = _mm_hsub_epi16(row2, row3);
row2 = _mm_hadd_epi16(row0, row1);
row3 = _mm_hsub_epi16(row0, row1);
//Abs and sum
row2 = _mm_abs_epi16(row2);
row3 = _mm_abs_epi16(row3);
row3 = _mm_add_epi16(row2, row3);
row3 = _mm_add_epi16(row3, _mm_shuffle_epi32(row3, KVZ_PERMUTE(2, 3, 0, 1) ));
row3 = _mm_add_epi16(row3, _mm_shuffle_epi32(row3, KVZ_PERMUTE(1, 0, 1, 0) ));
row3 = _mm_add_epi16(row3, _mm_shufflelo_epi16(row3, KVZ_PERMUTE(1, 0, 1, 0) ));
unsigned sum = _mm_extract_epi16(row3, 0);
unsigned satd = (sum + 1) >> 1;
return satd;
}
static INLINE void ver_add_sub_avx2(__m128i (*temp_hor)[8], __m128i (*temp_ver)[8]){
// First stage
for (int i = 0; i < 8; i += 2){
(*temp_ver)[i+0] = _mm_hadd_epi16((*temp_hor)[i + 0], (*temp_hor)[i + 1]);
(*temp_ver)[i+1] = _mm_hsub_epi16((*temp_hor)[i + 0], (*temp_hor)[i + 1]);
}
// Second stage
for (int i = 0; i < 8; i += 4){
(*temp_hor)[i + 0] = _mm_add_epi16((*temp_ver)[i + 0], (*temp_ver)[i + 2]);
(*temp_hor)[i + 1] = _mm_add_epi16((*temp_ver)[i + 1], (*temp_ver)[i + 3]);
(*temp_hor)[i + 2] = _mm_sub_epi16((*temp_ver)[i + 0], (*temp_ver)[i + 2]);
(*temp_hor)[i + 3] = _mm_sub_epi16((*temp_ver)[i + 1], (*temp_ver)[i + 3]);
}
// Third stage
for (int i = 0; i < 4; ++i){
(*temp_ver)[i + 0] = _mm_add_epi16((*temp_hor)[0 + i], (*temp_hor)[4 + i]);
(*temp_ver)[i + 4] = _mm_sub_epi16((*temp_hor)[0 + i], (*temp_hor)[4 + i]);
}
}
__m128i test_mm_hsub_epi16(__m128i a, __m128i b) {
// CHECK-LABEL: test_mm_hsub_epi16
// CHECK: call <8 x i16> @llvm.x86.ssse3.phsub.w.128(<8 x i16> %{{.*}}, <8 x i16> %{{.*}})
return _mm_hsub_epi16(a, b);
}
