hashirama/uvg266
1
0
Fork 0
mirror of https://github.com/ultravideo/uvg266.git synced 2024-12-04 13:54:05 +00:00
Code Issues Actions Projects Releases Wiki Activity

Change the right shift in pred_planar_avx2 to use a 128 bit register version of the right shift instrinsics, since when the integer version does not have a compile time constant the compiler is forced to generate the 128 bit register using version anyways, but also has to convert the integer to the 128 bit register, and the compiler does not optimize this properly and instead does the conversion on every call of the loop. ***THIS NEEDS TO BE DONE FOR ALL SHIFTS THAT DO NOT USE COMPILE TIME CONSTANT SHIFTS***

Browse source
This commit is contained in:
Joose Sainio 2023-09-12 14:07:30 +03:00
parent b02fb1b1af
commit 8bbf01c376
1 changed files with 4 additions and 2 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

6
src/strategies/avx2/intra-avx2.c
View file

@ -980,6 +980,8 @@ void uvg_intra_pred_planar_avx2(const cu_loc_t* const cu_loc,
}*/
// New loop
__m128i shift_r_v = _mm_setzero_si128();
shift_r_v = _mm_insert_epi32(shift_r_v, shift_r, 0);
for (int i = 0, d = 0; i < samples; i += 16, ++d) {
__m256i v_lo = _mm256_unpacklo_epi16(v_pred_hor[d], v_pred_ver[d]);
__m256i v_hi = _mm256_unpackhi_epi16(v_pred_hor[d], v_pred_ver[d]);
@ -991,8 +993,8 @@ void uvg_intra_pred_planar_avx2(const cu_loc_t* const cu_loc,
v_madd_lo = _mm256_add_epi32(v_madd_lo, v_samples);
v_madd_hi = _mm256_add_epi32(v_madd_hi, v_samples);
v_madd_lo = _mm256_srli_epi32(v_madd_lo, shift_r);
v_madd_hi = _mm256_srli_epi32(v_madd_hi, shift_r);
v_madd_lo = _mm256_srl_epi32(v_madd_lo, shift_r_v);
v_madd_hi = _mm256_srl_epi32(v_madd_hi, shift_r_v);
v_res[d] = _mm256_packs_epi32(v_madd_lo, v_madd_hi);
}