Skip to content

Commit

Permalink
deinterleave luma data
Browse files Browse the repository at this point in the history
  • Loading branch information
@wx257osn2
wx257osn2 committed Feb 5, 2024
1 parent fa6a9fd commit 4e8637c
Showing 1 changed file with 20 additions and 20 deletions.
40 changes: 20 additions & 20 deletions include/qoixx.hpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -516,12 +516,12 @@ class qoi{
hash = svand_n_u8_x(mask, svadd_u8_x(mask, svadd_u8_x(mask, svmul_n_u8_x(mask, get<0>(pxs), 3), svmul_n_u8_x(mask, get<1>(pxs), 5)), svadd_u8_x(mask, svmul_n_u8_x(mask, get<2>(pxs), 7), svmul_n_u8_x(mask, get<3>(pxs), 11))), 63);
else
hash = svand_n_u8_x(mask, svadd_u8_x(mask, svadd_u8_x(mask, svmul_n_u8_x(mask, get<0>(pxs), 3), svmul_n_u8_x(mask, get<1>(pxs), 5)), svadd_n_u8_x(mask, svmul_n_u8_x(mask, get<2>(pxs), 7), static_cast<std::uint8_t>(255*11))), 63);
std::uint8_t runs[SVERegisterSize/8], diffs[SVERegisterSize/8], lus[SVERegisterSize/8], mas[SVERegisterSize/8], hashs[SVERegisterSize/8];
std::uint8_t runs[SVERegisterSize/8], diffs[SVERegisterSize/8], lumas[SVERegisterSize/8*2], hashs[SVERegisterSize/8];
[[maybe_unused]] std::uint8_t alphas[SVERegisterSize/8];
svst1_u8(mask, runs, svadd_n_u8_m(runv, zero, 1));
svst1_u8(mask, diffs, diffv);
svst1_u8(mask, lus, lu);
svst1_u8(mask, mas, ma);
const auto luma = svcreate2_u8(lu, ma);
svst2_u8(mask, lumas, luma);
svst1_u8(mask, hashs, hash);
if constexpr(Alpha)
if(!alpha)
Expand Down Expand Up @@ -562,9 +562,9 @@ class qoi{
}
if(diffs[i])
*p++ = diffs[i];
else if(lus[i]){
*p++ = lus[i];
*p++ = mas[i];
else if(lumas[i*2]){
std::memcpy(p, lumas + i*2, 2);
p += 2;
}
else{
*p++ = chunk_tag::rgb;
Expand Down Expand Up @@ -683,12 +683,11 @@ class qoi{
hash = vandq_u8(vaddq_u8(vaddq_u8(vmulq_u8(pxs.val[0], vdupq_n_u8(3)), vmulq_u8(pxs.val[1], vdupq_n_u8(5))), vaddq_u8(vmulq_u8(pxs.val[2], vdupq_n_u8(7)), vmulq_u8(pxs.val[3], vdupq_n_u8(11)))), vdupq_n_u8(63));
else
hash = vandq_u8(vaddq_u8(vaddq_u8(vmulq_u8(pxs.val[0], vdupq_n_u8(3)), vmulq_u8(pxs.val[1], vdupq_n_u8(5))), vaddq_u8(vmulq_u8(pxs.val[2], vdupq_n_u8(7)), vdupq_n_u8(static_cast<std::uint8_t>(255*11)))), vdupq_n_u8(63));
std::uint8_t runs[simd_lanes], diffs[simd_lanes], lus[simd_lanes], mas[simd_lanes], hashs[simd_lanes];
std::uint8_t runs[simd_lanes], diffs[simd_lanes], lumas[simd_lanes*2], hashs[simd_lanes];
[[maybe_unused]] std::uint8_t alphas[simd_lanes];
vst1q_u8(runs, runv);
vst1q_u8(diffs, diffv);
vst1q_u8(lus, lu);
vst1q_u8(mas, ma);
vst2q_u8(lumas, uint8x16x2_t{lu, ma});
vst1q_u8(hashs, hash);
if constexpr(Alpha)
if(!alpha)
Expand Down Expand Up @@ -729,9 +728,9 @@ class qoi{
}
if(diffs[i])
*p++ = diffs[i];
else if(lus[i]){
*p++ = lus[i];
*p++ = mas[i];
else if(lumas[i*2]){
std::memcpy(p, lumas + i*2, 2);
p += 2;
}
else{
*p++ = chunk_tag::rgb;
Expand Down Expand Up @@ -954,19 +953,20 @@ class qoi{
diff.val[0] = _mm256_add_epi8(_mm256_sub_epi8(diff.val[0], diff.val[1]), eight);
diff.val[2] = _mm256_add_epi8(_mm256_sub_epi8(diff.val[2], diff.val[1]), eight);
diff.val[1] = _mm256_add_epi8(diff.val[1], _mm256_set1_epi8(30));
const auto lu = _mm256_and_si256(_mm256_or_si256(_mm256_set1_epi8(static_cast<char>(chunk_tag::luma)), diff.val[1]), _mm256_cmpeq_epi8(_mm256_or_si256(_mm256_and_si256(_mm256_or_si256(diff.val[0], diff.val[2]), _mm256_set1_epi8(static_cast<char>(0xf0))), _mm256_and_si256(diff.val[1], _mm256_set1_epi8(static_cast<char>(0xc0)))), zero));
const auto ma = _mm256_or_si256(slli_epi8<4>(diff.val[0]), diff.val[2]);
const auto luma_mask = _mm256_setr_epi32(0, 1, 4, 5, 2, 3, 6, 7);
const auto lu = _mm256_permutevar8x32_epi32(_mm256_and_si256(_mm256_or_si256(_mm256_set1_epi8(static_cast<char>(chunk_tag::luma)), diff.val[1]), _mm256_cmpeq_epi8(_mm256_or_si256(_mm256_and_si256(_mm256_or_si256(diff.val[0], diff.val[2]), _mm256_set1_epi8(static_cast<char>(0xf0))), _mm256_and_si256(diff.val[1], _mm256_set1_epi8(static_cast<char>(0xc0)))), zero)), luma_mask);
const auto ma = _mm256_permutevar8x32_epi32(_mm256_or_si256(slli_epi8<4>(diff.val[0]), diff.val[2]), luma_mask);
__m256i hash;
if constexpr(Alpha)
hash = _mm256_and_si256(_mm256_add_epi8(_mm256_add_epi8(mul_epi8<3>(pxs.val[0]), mul_epi8<5>(pxs.val[1])), _mm256_add_epi8(mul_epi8<7>(pxs.val[2]), mul_epi8<11>(pxs.val[3]))), _mm256_set1_epi8(63));
else
hash = _mm256_and_si256(_mm256_add_epi8(_mm256_add_epi8(mul_epi8<3>(pxs.val[0]), mul_epi8<5>(pxs.val[1])), _mm256_add_epi8(mul_epi8<7>(pxs.val[2]), _mm256_set1_epi8(static_cast<std::uint8_t>(255*11)))), _mm256_set1_epi8(63));
alignas(alignof(__m256i)) std::uint8_t runs[simd_lanes], diffs[simd_lanes], lus[simd_lanes], mas[simd_lanes], hashs[simd_lanes];
alignas(alignof(__m256i)) std::uint8_t runs[simd_lanes], diffs[simd_lanes], lumas[simd_lanes*2], hashs[simd_lanes];
[[maybe_unused]] alignas(alignof(__m256i)) std::uint8_t alphas[simd_lanes];
_mm256_store_si256(reinterpret_cast<__m256i*>(runs), runv);
_mm256_store_si256(reinterpret_cast<__m256i*>(diffs), diffv);
_mm256_store_si256(reinterpret_cast<__m256i*>(lus), lu);
_mm256_store_si256(reinterpret_cast<__m256i*>(mas), ma);
_mm256_store_si256(reinterpret_cast<__m256i*>(lumas), _mm256_unpacklo_epi8(lu, ma));
_mm256_store_si256(reinterpret_cast<__m256i*>(lumas)+1, _mm256_unpackhi_epi8(lu, ma));
_mm256_store_si256(reinterpret_cast<__m256i*>(hashs), hash);
if constexpr(Alpha)
if(!alpha)
Expand Down Expand Up @@ -1007,9 +1007,9 @@ class qoi{
}
if(diffs[i])
*p++ = diffs[i];
else if(lus[i]){
*p++ = lus[i];
*p++ = mas[i];
else if(lumas[i*2]){
std::memcpy(p, lumas + i*2, 2);
p += 2;
}
else{
*p++ = chunk_tag::rgb;
Expand Down

0 comments on commit 4e8637c

Please sign in to comment.