pixconv: add swizzle_ycc__convert_rgbx_x86_avx2

Some of the benchmark numbers move around even though they don't use the
new etc__convert_rgbx_x86_avx2 code paths. It's possibly just a code
alignment coincidence per https://news.ycombinator.com/item?id=23633583

name                                             old speed     new speed     delta

wuffs_jpeg_decode_19k_8bpp/clang11               372MB/s ± 0%  370MB/s ± 0%  -0.64%  (p=0.008 n=5+5)
wuffs_jpeg_decode_30k_24bpp_progressive/clang11  278MB/s ± 0%  295MB/s ± 0%  +5.89%  (p=0.008 n=5+5)
wuffs_jpeg_decode_30k_24bpp_sequential/clang11   710MB/s ± 0%  726MB/s ± 0%  +2.30%  (p=0.008 n=5+5)
wuffs_jpeg_decode_77k_24bpp/clang11              603MB/s ± 0%  622MB/s ± 0%  +3.12%  (p=0.008 n=5+5)
wuffs_jpeg_decode_552k_24bpp_420/clang11         635MB/s ± 0%  659MB/s ± 0%  +3.82%  (p=0.008 n=5+5)
wuffs_jpeg_decode_552k_24bpp_444/clang11         650MB/s ± 0%  649MB/s ± 0%  -0.17%  (p=0.008 n=5+5)
wuffs_jpeg_decode_4002k_24bpp/clang11            629MB/s ± 0%  675MB/s ± 0%  +7.20%  (p=0.008 n=5+5)

wuffs_jpeg_decode_19k_8bpp/gcc10                 364MB/s ± 0%  351MB/s ± 0%  -3.62%  (p=0.008 n=5+5)
wuffs_jpeg_decode_30k_24bpp_progressive/gcc10    334MB/s ± 0%  334MB/s ± 0%  -0.06%  (p=0.032 n=5+5)
wuffs_jpeg_decode_30k_24bpp_sequential/gcc10     944MB/s ± 0%  924MB/s ± 0%  -2.19%  (p=0.008 n=5+5)
wuffs_jpeg_decode_77k_24bpp/gcc10                769MB/s ± 0%  740MB/s ± 0%  -3.75%  (p=0.008 n=5+5)
wuffs_jpeg_decode_552k_24bpp_420/gcc10           842MB/s ± 0%  812MB/s ± 0%  -3.60%  (p=0.008 n=5+5)
wuffs_jpeg_decode_552k_24bpp_444/gcc10           661MB/s ± 0%  647MB/s ± 0%  -2.13%  (p=0.008 n=5+5)
wuffs_jpeg_decode_4002k_24bpp/gcc10              849MB/s ± 0%  817MB/s ± 0%  -3.77%  (p=0.016 n=5+4)
This commit is contained in:
Nigel Tao committed 2023-07-22 13:11:12 +10:00
1 parent 4facef4e01
commit 35c1f76f86
3 files changed
+338

No files matched your search

@@ -222,6 +222,8 @@ wuffs_base__pixel_swizzler__swizzle_ycc__convert_bgrx_x86_avx2(
// mix01 = [b01 g01 b03 g03 .. b0F g0F b11 g11 .. b1D g1D b1F g1F]
// mix02 = [r00 % r02 % .. r0E % r10 % .. r1C % r1E %]
// mix03 = [r01 % r03 % .. r0F % r11 % .. r1D % r1F %]
//
// See also § below.
__m256i mix00 = _mm256_unpacklo_epi8(packed_b_eve, packed_g_eve);
__m256i mix01 = _mm256_unpacklo_epi8(packed_b_odd, packed_g_odd);
__m256i mix02 = _mm256_unpacklo_epi8(packed_r_eve, uFFFF);
@@ -282,5 +284,154 @@ wuffs_base__pixel_swizzler__swizzle_ycc__convert_bgrx_x86_avx2(
x += n;
}
}
// The rgbx flavor (below) is exactly the same as the bgrx flavor (above)
// except for the lines marked with a § and that comments were stripped.
WUFFS_BASE__MAYBE_ATTRIBUTE_TARGET("pclmul,popcnt,sse4.2,avx2")
static void //
wuffs_base__pixel_swizzler__swizzle_ycc__convert_rgbx_x86_avx2(
wuffs_base__pixel_buffer* dst,
uint32_t x,
uint32_t x_end,
uint32_t y,
const uint8_t* up0,
const uint8_t* up1,
const uint8_t* up2) {
if ((x + 32u) > x_end) {
wuffs_base__pixel_swizzler__swizzle_ycc__convert_bgrx( //
dst, x, x_end, y, up0, up1, up2);
return;
}
size_t dst_stride = dst->private_impl.planes[0].stride;
uint8_t* dst_iter = dst->private_impl.planes[0].ptr +
(dst_stride * ((size_t)y)) + (4 * ((size_t)x));
const __m256i u0001 = _mm256_set1_epi16(0x0001);
const __m256i u00FF = _mm256_set1_epi16(0x00FF);
const __m256i uFF80 = _mm256_set1_epi16(0xFF80);
const __m256i uFFFF = _mm256_set1_epi16(0xFFFF);
const __m256i p8000_p0000 = _mm256_set_epi16( //
0x0000, 0x8000, 0x0000, 0x8000, 0x0000, 0x8000, 0x0000, 0x8000, //
0x0000, 0x8000, 0x0000, 0x8000, 0x0000, 0x8000, 0x0000, 0x8000);
const __m256i m3A5E = _mm256_set1_epi16(-0x3A5E);
const __m256i p66E9 = _mm256_set1_epi16(+0x66E9);
const __m256i m581A_p492E = _mm256_set_epi16( //
+0x492E, -0x581A, +0x492E, -0x581A, //
+0x492E, -0x581A, +0x492E, -0x581A, //
+0x492E, -0x581A, +0x492E, -0x581A, //
+0x492E, -0x581A, +0x492E, -0x581A);
while (x < x_end) {
__m256i cb_all = _mm256_lddqu_si256((const __m256i*)(const void*)up1);
__m256i cr_all = _mm256_lddqu_si256((const __m256i*)(const void*)up2);
__m256i cb_eve = _mm256_add_epi16(uFF80, _mm256_and_si256(cb_all, u00FF));
__m256i cr_eve = _mm256_add_epi16(uFF80, _mm256_and_si256(cr_all, u00FF));
__m256i cb_odd = _mm256_add_epi16(uFF80, _mm256_srli_epi16(cb_all, 8));
__m256i cr_odd = _mm256_add_epi16(uFF80, _mm256_srli_epi16(cr_all, 8));
__m256i tmp_by_eve = _mm256_srai_epi16(
_mm256_add_epi16(
_mm256_mulhi_epi16(_mm256_add_epi16(cb_eve, cb_eve), m3A5E), u0001),
1);
__m256i tmp_by_odd = _mm256_srai_epi16(
_mm256_add_epi16(
_mm256_mulhi_epi16(_mm256_add_epi16(cb_odd, cb_odd), m3A5E), u0001),
1);
__m256i tmp_ry_eve = _mm256_srai_epi16(
_mm256_add_epi16(
_mm256_mulhi_epi16(_mm256_add_epi16(cr_eve, cr_eve), p66E9), u0001),
1);
__m256i tmp_ry_odd = _mm256_srai_epi16(
_mm256_add_epi16(
_mm256_mulhi_epi16(_mm256_add_epi16(cr_odd, cr_odd), p66E9), u0001),
1);
__m256i by_eve =
_mm256_add_epi16(tmp_by_eve, _mm256_add_epi16(cb_eve, cb_eve));
__m256i by_odd =
_mm256_add_epi16(tmp_by_odd, _mm256_add_epi16(cb_odd, cb_odd));
__m256i ry_eve = _mm256_add_epi16(tmp_ry_eve, cr_eve);
__m256i ry_odd = _mm256_add_epi16(tmp_ry_odd, cr_odd);
__m256i tmp0_gy_eve_lo = _mm256_madd_epi16( //
_mm256_unpacklo_epi16(cb_eve, cr_eve), m581A_p492E);
__m256i tmp0_gy_eve_hi = _mm256_madd_epi16( //
_mm256_unpackhi_epi16(cb_eve, cr_eve), m581A_p492E);
__m256i tmp0_gy_odd_lo = _mm256_madd_epi16( //
_mm256_unpacklo_epi16(cb_odd, cr_odd), m581A_p492E);
__m256i tmp0_gy_odd_hi = _mm256_madd_epi16( //
_mm256_unpackhi_epi16(cb_odd, cr_odd), m581A_p492E);
__m256i tmp1_gy_eve_lo =
_mm256_srai_epi32(_mm256_add_epi32(tmp0_gy_eve_lo, p8000_p0000), 16);
__m256i tmp1_gy_eve_hi =
_mm256_srai_epi32(_mm256_add_epi32(tmp0_gy_eve_hi, p8000_p0000), 16);
__m256i tmp1_gy_odd_lo =
_mm256_srai_epi32(_mm256_add_epi32(tmp0_gy_odd_lo, p8000_p0000), 16);
__m256i tmp1_gy_odd_hi =
_mm256_srai_epi32(_mm256_add_epi32(tmp0_gy_odd_hi, p8000_p0000), 16);
__m256i gy_eve = _mm256_sub_epi16(
_mm256_packs_epi32(tmp1_gy_eve_lo, tmp1_gy_eve_hi), cr_eve);
__m256i gy_odd = _mm256_sub_epi16(
_mm256_packs_epi32(tmp1_gy_odd_lo, tmp1_gy_odd_hi), cr_odd);
__m256i yy_all = _mm256_lddqu_si256((const __m256i*)(const void*)up0);
__m256i yy_eve = _mm256_and_si256(yy_all, u00FF);
__m256i yy_odd = _mm256_srli_epi16(yy_all, 8);
__m256i loose_b_eve = _mm256_add_epi16(by_eve, yy_eve);
__m256i loose_b_odd = _mm256_add_epi16(by_odd, yy_odd);
__m256i packed_b_eve = _mm256_packus_epi16(loose_b_eve, loose_b_eve);
__m256i packed_b_odd = _mm256_packus_epi16(loose_b_odd, loose_b_odd);
__m256i loose_g_eve = _mm256_add_epi16(gy_eve, yy_eve);
__m256i loose_g_odd = _mm256_add_epi16(gy_odd, yy_odd);
__m256i packed_g_eve = _mm256_packus_epi16(loose_g_eve, loose_g_eve);
__m256i packed_g_odd = _mm256_packus_epi16(loose_g_odd, loose_g_odd);
__m256i loose_r_eve = _mm256_add_epi16(ry_eve, yy_eve);
__m256i loose_r_odd = _mm256_add_epi16(ry_odd, yy_odd);
__m256i packed_r_eve = _mm256_packus_epi16(loose_r_eve, loose_r_eve);
__m256i packed_r_odd = _mm256_packus_epi16(loose_r_odd, loose_r_odd);
// § Note the swapped B and R channels.
__m256i mix00 = _mm256_unpacklo_epi8(packed_r_eve, packed_g_eve);
__m256i mix01 = _mm256_unpacklo_epi8(packed_r_odd, packed_g_odd);
__m256i mix02 = _mm256_unpacklo_epi8(packed_b_eve, uFFFF);
__m256i mix03 = _mm256_unpacklo_epi8(packed_b_odd, uFFFF);
__m256i mix10 = _mm256_unpacklo_epi16(mix00, mix02);
__m256i mix11 = _mm256_unpacklo_epi16(mix01, mix03);
__m256i mix12 = _mm256_unpackhi_epi16(mix00, mix02);
__m256i mix13 = _mm256_unpackhi_epi16(mix01, mix03);
__m256i mix20 = _mm256_unpacklo_epi32(mix10, mix11);
__m256i mix21 = _mm256_unpackhi_epi32(mix10, mix11);
__m256i mix22 = _mm256_unpacklo_epi32(mix12, mix13);
__m256i mix23 = _mm256_unpackhi_epi32(mix12, mix13);
__m256i mix30 = _mm256_permute2x128_si256(mix20, mix21, 0x20);
__m256i mix31 = _mm256_permute2x128_si256(mix22, mix23, 0x20);
__m256i mix32 = _mm256_permute2x128_si256(mix20, mix21, 0x31);
__m256i mix33 = _mm256_permute2x128_si256(mix22, mix23, 0x31);
_mm256_storeu_si256((__m256i*)(void*)(dst_iter + 0x00), mix30);
_mm256_storeu_si256((__m256i*)(void*)(dst_iter + 0x20), mix31);
_mm256_storeu_si256((__m256i*)(void*)(dst_iter + 0x40), mix32);
_mm256_storeu_si256((__m256i*)(void*)(dst_iter + 0x60), mix33);
uint32_t n = 32 - (31 & (x - x_end));
dst_iter += 4 * n;
up0 += n;
up1 += n;
up2 += n;
x += n;
}
}
#endif // defined(WUFFS_BASE__CPU_ARCH__X86_FAMILY)
// ‼ WUFFS MULTI-FILE SECTION -x86_avx2
@@ -25,6 +25,17 @@ wuffs_base__pixel_swizzler__swizzle_ycc__convert_bgrx_x86_avx2(
const uint8_t* up0,
const uint8_t* up1,
const uint8_t* up2);
WUFFS_BASE__MAYBE_ATTRIBUTE_TARGET("pclmul,popcnt,sse4.2,avx2")
static void //
wuffs_base__pixel_swizzler__swizzle_ycc__convert_rgbx_x86_avx2(
wuffs_base__pixel_buffer* dst,
uint32_t x,
uint32_t x_end,
uint32_t y,
const uint8_t* up0,
const uint8_t* up1,
const uint8_t* up2);
#endif
// --------
@@ -977,6 +988,13 @@ wuffs_base__pixel_swizzler__swizzle_ycck(
case WUFFS_BASE__PIXEL_FORMAT__RGBA_NONPREMUL:
case WUFFS_BASE__PIXEL_FORMAT__RGBA_PREMUL:
case WUFFS_BASE__PIXEL_FORMAT__RGBX:
#if defined(WUFFS_BASE__CPU_ARCH__X86_FAMILY)
if (wuffs_base__cpu_arch__have_x86_sse42()) {
convfunc =
&wuffs_base__pixel_swizzler__swizzle_ycc__convert_rgbx_x86_avx2;
break;
}
#endif
convfunc = &wuffs_base__pixel_swizzler__swizzle_ycc__convert_rgbx;
break;
default:
+169
View File
@@ -23526,6 +23526,17 @@ wuffs_base__pixel_swizzler__swizzle_ycc__convert_bgrx_x86_avx2(
const uint8_t* up0,
const uint8_t* up1,
const uint8_t* up2);
WUFFS_BASE__MAYBE_ATTRIBUTE_TARGET("pclmul,popcnt,sse4.2,avx2")
static void //
wuffs_base__pixel_swizzler__swizzle_ycc__convert_rgbx_x86_avx2(
wuffs_base__pixel_buffer* dst,
uint32_t x,
uint32_t x_end,
uint32_t y,
const uint8_t* up0,
const uint8_t* up1,
const uint8_t* up2);
#endif
// --------
@@ -24742,6 +24753,13 @@ wuffs_base__pixel_swizzler__swizzle_ycck(
case WUFFS_BASE__PIXEL_FORMAT__RGBA_NONPREMUL:
case WUFFS_BASE__PIXEL_FORMAT__RGBA_PREMUL:
case WUFFS_BASE__PIXEL_FORMAT__RGBX:
#if defined(WUFFS_BASE__CPU_ARCH__X86_FAMILY)
if (wuffs_base__cpu_arch__have_x86_sse42()) {
convfunc =
&wuffs_base__pixel_swizzler__swizzle_ycc__convert_rgbx_x86_avx2;
break;
}
#endif
convfunc = &wuffs_base__pixel_swizzler__swizzle_ycc__convert_rgbx;
break;
default:
@@ -25020,6 +25038,8 @@ wuffs_base__pixel_swizzler__swizzle_ycc__convert_bgrx_x86_avx2(
// mix01 = [b01 g01 b03 g03 .. b0F g0F b11 g11 .. b1D g1D b1F g1F]
// mix02 = [r00 % r02 % .. r0E % r10 % .. r1C % r1E %]
// mix03 = [r01 % r03 % .. r0F % r11 % .. r1D % r1F %]
//
// See also § below.
__m256i mix00 = _mm256_unpacklo_epi8(packed_b_eve, packed_g_eve);
__m256i mix01 = _mm256_unpacklo_epi8(packed_b_odd, packed_g_odd);
__m256i mix02 = _mm256_unpacklo_epi8(packed_r_eve, uFFFF);
@@ -25080,6 +25100,155 @@ wuffs_base__pixel_swizzler__swizzle_ycc__convert_bgrx_x86_avx2(
x += n;
}
}
// The rgbx flavor (below) is exactly the same as the bgrx flavor (above)
// except for the lines marked with a § and that comments were stripped.
WUFFS_BASE__MAYBE_ATTRIBUTE_TARGET("pclmul,popcnt,sse4.2,avx2")
static void //
wuffs_base__pixel_swizzler__swizzle_ycc__convert_rgbx_x86_avx2(
wuffs_base__pixel_buffer* dst,
uint32_t x,
uint32_t x_end,
uint32_t y,
const uint8_t* up0,
const uint8_t* up1,
const uint8_t* up2) {
if ((x + 32u) > x_end) {
wuffs_base__pixel_swizzler__swizzle_ycc__convert_bgrx( //
dst, x, x_end, y, up0, up1, up2);
return;
}
size_t dst_stride = dst->private_impl.planes[0].stride;
uint8_t* dst_iter = dst->private_impl.planes[0].ptr +
(dst_stride * ((size_t)y)) + (4 * ((size_t)x));
const __m256i u0001 = _mm256_set1_epi16(0x0001);
const __m256i u00FF = _mm256_set1_epi16(0x00FF);
const __m256i uFF80 = _mm256_set1_epi16(0xFF80);
const __m256i uFFFF = _mm256_set1_epi16(0xFFFF);
const __m256i p8000_p0000 = _mm256_set_epi16( //
0x0000, 0x8000, 0x0000, 0x8000, 0x0000, 0x8000, 0x0000, 0x8000, //
0x0000, 0x8000, 0x0000, 0x8000, 0x0000, 0x8000, 0x0000, 0x8000);
const __m256i m3A5E = _mm256_set1_epi16(-0x3A5E);
const __m256i p66E9 = _mm256_set1_epi16(+0x66E9);
const __m256i m581A_p492E = _mm256_set_epi16( //
+0x492E, -0x581A, +0x492E, -0x581A, //
+0x492E, -0x581A, +0x492E, -0x581A, //
+0x492E, -0x581A, +0x492E, -0x581A, //
+0x492E, -0x581A, +0x492E, -0x581A);
while (x < x_end) {
__m256i cb_all = _mm256_lddqu_si256((const __m256i*)(const void*)up1);
__m256i cr_all = _mm256_lddqu_si256((const __m256i*)(const void*)up2);
__m256i cb_eve = _mm256_add_epi16(uFF80, _mm256_and_si256(cb_all, u00FF));
__m256i cr_eve = _mm256_add_epi16(uFF80, _mm256_and_si256(cr_all, u00FF));
__m256i cb_odd = _mm256_add_epi16(uFF80, _mm256_srli_epi16(cb_all, 8));
__m256i cr_odd = _mm256_add_epi16(uFF80, _mm256_srli_epi16(cr_all, 8));
__m256i tmp_by_eve = _mm256_srai_epi16(
_mm256_add_epi16(
_mm256_mulhi_epi16(_mm256_add_epi16(cb_eve, cb_eve), m3A5E), u0001),
1);
__m256i tmp_by_odd = _mm256_srai_epi16(
_mm256_add_epi16(
_mm256_mulhi_epi16(_mm256_add_epi16(cb_odd, cb_odd), m3A5E), u0001),
1);
__m256i tmp_ry_eve = _mm256_srai_epi16(
_mm256_add_epi16(
_mm256_mulhi_epi16(_mm256_add_epi16(cr_eve, cr_eve), p66E9), u0001),
1);
__m256i tmp_ry_odd = _mm256_srai_epi16(
_mm256_add_epi16(
_mm256_mulhi_epi16(_mm256_add_epi16(cr_odd, cr_odd), p66E9), u0001),
1);
__m256i by_eve =
_mm256_add_epi16(tmp_by_eve, _mm256_add_epi16(cb_eve, cb_eve));
__m256i by_odd =
_mm256_add_epi16(tmp_by_odd, _mm256_add_epi16(cb_odd, cb_odd));
__m256i ry_eve = _mm256_add_epi16(tmp_ry_eve, cr_eve);
__m256i ry_odd = _mm256_add_epi16(tmp_ry_odd, cr_odd);
__m256i tmp0_gy_eve_lo = _mm256_madd_epi16( //
_mm256_unpacklo_epi16(cb_eve, cr_eve), m581A_p492E);
__m256i tmp0_gy_eve_hi = _mm256_madd_epi16( //
_mm256_unpackhi_epi16(cb_eve, cr_eve), m581A_p492E);
__m256i tmp0_gy_odd_lo = _mm256_madd_epi16( //
_mm256_unpacklo_epi16(cb_odd, cr_odd), m581A_p492E);
__m256i tmp0_gy_odd_hi = _mm256_madd_epi16( //
_mm256_unpackhi_epi16(cb_odd, cr_odd), m581A_p492E);
__m256i tmp1_gy_eve_lo =
_mm256_srai_epi32(_mm256_add_epi32(tmp0_gy_eve_lo, p8000_p0000), 16);
__m256i tmp1_gy_eve_hi =
_mm256_srai_epi32(_mm256_add_epi32(tmp0_gy_eve_hi, p8000_p0000), 16);
__m256i tmp1_gy_odd_lo =
_mm256_srai_epi32(_mm256_add_epi32(tmp0_gy_odd_lo, p8000_p0000), 16);
__m256i tmp1_gy_odd_hi =
_mm256_srai_epi32(_mm256_add_epi32(tmp0_gy_odd_hi, p8000_p0000), 16);
__m256i gy_eve = _mm256_sub_epi16(
_mm256_packs_epi32(tmp1_gy_eve_lo, tmp1_gy_eve_hi), cr_eve);
__m256i gy_odd = _mm256_sub_epi16(
_mm256_packs_epi32(tmp1_gy_odd_lo, tmp1_gy_odd_hi), cr_odd);
__m256i yy_all = _mm256_lddqu_si256((const __m256i*)(const void*)up0);
__m256i yy_eve = _mm256_and_si256(yy_all, u00FF);
__m256i yy_odd = _mm256_srli_epi16(yy_all, 8);
__m256i loose_b_eve = _mm256_add_epi16(by_eve, yy_eve);
__m256i loose_b_odd = _mm256_add_epi16(by_odd, yy_odd);
__m256i packed_b_eve = _mm256_packus_epi16(loose_b_eve, loose_b_eve);
__m256i packed_b_odd = _mm256_packus_epi16(loose_b_odd, loose_b_odd);
__m256i loose_g_eve = _mm256_add_epi16(gy_eve, yy_eve);
__m256i loose_g_odd = _mm256_add_epi16(gy_odd, yy_odd);
__m256i packed_g_eve = _mm256_packus_epi16(loose_g_eve, loose_g_eve);
__m256i packed_g_odd = _mm256_packus_epi16(loose_g_odd, loose_g_odd);
__m256i loose_r_eve = _mm256_add_epi16(ry_eve, yy_eve);
__m256i loose_r_odd = _mm256_add_epi16(ry_odd, yy_odd);
__m256i packed_r_eve = _mm256_packus_epi16(loose_r_eve, loose_r_eve);
__m256i packed_r_odd = _mm256_packus_epi16(loose_r_odd, loose_r_odd);
// § Note the swapped B and R channels.
__m256i mix00 = _mm256_unpacklo_epi8(packed_r_eve, packed_g_eve);
__m256i mix01 = _mm256_unpacklo_epi8(packed_r_odd, packed_g_odd);
__m256i mix02 = _mm256_unpacklo_epi8(packed_b_eve, uFFFF);
__m256i mix03 = _mm256_unpacklo_epi8(packed_b_odd, uFFFF);
__m256i mix10 = _mm256_unpacklo_epi16(mix00, mix02);
__m256i mix11 = _mm256_unpacklo_epi16(mix01, mix03);
__m256i mix12 = _mm256_unpackhi_epi16(mix00, mix02);
__m256i mix13 = _mm256_unpackhi_epi16(mix01, mix03);
__m256i mix20 = _mm256_unpacklo_epi32(mix10, mix11);
__m256i mix21 = _mm256_unpackhi_epi32(mix10, mix11);
__m256i mix22 = _mm256_unpacklo_epi32(mix12, mix13);
__m256i mix23 = _mm256_unpackhi_epi32(mix12, mix13);
__m256i mix30 = _mm256_permute2x128_si256(mix20, mix21, 0x20);
__m256i mix31 = _mm256_permute2x128_si256(mix22, mix23, 0x20);
__m256i mix32 = _mm256_permute2x128_si256(mix20, mix21, 0x31);
__m256i mix33 = _mm256_permute2x128_si256(mix22, mix23, 0x31);
_mm256_storeu_si256((__m256i*)(void*)(dst_iter + 0x00), mix30);
_mm256_storeu_si256((__m256i*)(void*)(dst_iter + 0x20), mix31);
_mm256_storeu_si256((__m256i*)(void*)(dst_iter + 0x40), mix32);
_mm256_storeu_si256((__m256i*)(void*)(dst_iter + 0x60), mix33);
uint32_t n = 32 - (31 & (x - x_end));
dst_iter += 4 * n;
up0 += n;
up1 += n;
up2 += n;
x += n;
}
}
#endif // defined(WUFFS_BASE__CPU_ARCH__X86_FAMILY)
// ‼ WUFFS MULTI-FILE SECTION -x86_avx2