hashirama/uvg266
1
0
Fork 0
mirror of https://github.com/ultravideo/uvg266.git synced 2024-11-27 19:24:06 +00:00
Code Issues Actions Projects Releases Wiki Activity

Optimize rearrange_512 function

Browse source
This commit is contained in:
Pauli Oikkonen 2018-11-29 15:22:34 +02:00
parent cb8209d1b3
commit 6bbd3e5a44
1 changed files with 4 additions and 6 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

10
src/strategies/avx2/quant-avx2.c
View file

@ -88,13 +88,10 @@ static INLINE void get_first_last_nz_int16(__m256i ints, int32_t *first, int32_t
// (abcd|efgh) (ijkl|mnop) => (aceg|ikmo) (bdfh|jlnp) // (abcd|efgh) (ijkl|mnop) => (aceg|ikmo) (bdfh|jlnp)
static INLINE void rearrange_512(__m256i *hi, __m256i *lo) static INLINE void rearrange_512(__m256i *hi, __m256i *lo)
{ {
__m256i tmphi, tmplo; const __m256i perm8x32mask = _mm256_setr_epi32(0, 2, 4, 6, 1, 3, 5, 7);
tmphi = _mm256_shuffle_epi32(*hi, _MM_SHUFFLE(3, 1, 2, 0)); __m256i tmphi = _mm256_permutevar8x32_epi32(*hi, perm8x32mask);
tmplo = _mm256_shuffle_epi32(*lo, _MM_SHUFFLE(3, 1, 2, 0)); __m256i tmplo = _mm256_permutevar8x32_epi32(*lo, perm8x32mask);
tmphi = _mm256_permute4x64_epi64(tmphi, _MM_SHUFFLE(3, 1, 2, 0));
tmplo = _mm256_permute4x64_epi64(tmplo, _MM_SHUFFLE(3, 1, 2, 0));
*hi = _mm256_permute2x128_si256(tmplo, tmphi, 0x31); *hi = _mm256_permute2x128_si256(tmplo, tmphi, 0x31);
*lo = _mm256_permute2x128_si256(tmplo, tmphi, 0x20); *lo = _mm256_permute2x128_si256(tmplo, tmphi, 0x20);
@ -116,6 +113,7 @@ static INLINE void get_cheapest_alternative(__m256i costs_hi, __m256i costs_lo,
// Interleave ns and lo into 32-bit variables and to two 256-bit wide vecs, // Interleave ns and lo into 32-bit variables and to two 256-bit wide vecs,
// to have the same data layout as in costs. Zero extend to 32b width, shift // to have the same data layout as in costs. Zero extend to 32b width, shift
// changes 16 bits to the left, and store them into the same vectors. // changes 16 bits to the left, and store them into the same vectors.
// TODO: unpack instead of this
tmp1 = _mm256_cvtepu16_epi32(nslo); tmp1 = _mm256_cvtepu16_epi32(nslo);
tmp2 = _mm256_cvtepu16_epi32(chlo); tmp2 = _mm256_cvtepu16_epi32(chlo);
tmp2 = _mm256_bslli_epi128(tmp2, 2); tmp2 = _mm256_bslli_epi128(tmp2, 2);