static inline __m128i hardlight_byte_SSE2(const __m128i& sc, const __m128i& dc,
const __m128i& sa, const __m128i& da) {
// if (2 * sc <= sa)
__m128i tmp1 = _mm_slli_epi32(sc, 1);
__m128i cmp1 = _mm_cmpgt_epi32(tmp1, sa);
__m128i rc1 = _mm_mullo_epi16(sc, dc); // sc * dc;
rc1 = _mm_slli_epi32(rc1, 1); // 2 * sc * dc
rc1 = _mm_andnot_si128(cmp1, rc1);
// else
tmp1 = _mm_mullo_epi16(sa, da);
__m128i tmp2 = Multiply32_SSE2(_mm_sub_epi32(da, dc),
_mm_sub_epi32(sa, sc));
tmp2 = _mm_slli_epi32(tmp2, 1);
__m128i rc2 = _mm_sub_epi32(tmp1, tmp2);
rc2 = _mm_and_si128(cmp1, rc2);
__m128i rc = _mm_or_si128(rc1, rc2);
__m128i ida = _mm_sub_epi32(_mm_set1_epi32(255), da);
tmp1 = _mm_mullo_epi16(sc, ida);
__m128i isa = _mm_sub_epi32(_mm_set1_epi32(255), sa);
tmp2 = _mm_mullo_epi16(dc, isa);
rc = _mm_add_epi32(rc, tmp1);
rc = _mm_add_epi32(rc, tmp2);
return clamp_div255round_SSE2(rc);
}
static inline __m128i exclusion_byte_SSE2(const __m128i& sc, const __m128i& dc,
const __m128i&, __m128i&) {
__m128i tmp1 = _mm_mullo_epi16(_mm_set1_epi32(255), sc); // 255 * sc
__m128i tmp2 = _mm_mullo_epi16(_mm_set1_epi32(255), dc); // 255 * dc
tmp1 = _mm_add_epi32(tmp1, tmp2);
tmp2 = _mm_mullo_epi16(sc, dc); // sc * dc
tmp2 = _mm_slli_epi32(tmp2, 1); // 2 * sc * dc
__m128i r = _mm_sub_epi32(tmp1, tmp2);
return clamp_div255round_SSE2(r);
}
static inline __m128i colorburn_byte_SSE2(const __m128i& sc, const __m128i& dc,
const __m128i& sa, const __m128i& da) {
__m128i ida = _mm_sub_epi32(_mm_set1_epi32(255), da);
__m128i isa = _mm_sub_epi32(_mm_set1_epi32(255), sa);
// if (dc == da)
__m128i cmp1 = _mm_cmpeq_epi32(dc, da);
__m128i tmp1 = _mm_mullo_epi16(sa, da);
__m128i tmp2 = _mm_mullo_epi16(sc, ida);
__m128i tmp3 = _mm_mullo_epi16(dc, isa);
__m128i rc1 = _mm_add_epi32(tmp1, tmp2);
rc1 = _mm_add_epi32(rc1, tmp3);
rc1 = clamp_div255round_SSE2(rc1);
rc1 = _mm_and_si128(cmp1, rc1);
// else if (0 == sc)
__m128i cmp2 = _mm_cmpeq_epi32(sc, _mm_setzero_si128());
__m128i rc2 = SkAlphaMulAlpha_SSE2(dc, isa);
__m128i cmp = _mm_andnot_si128(cmp1, cmp2);
rc2 = _mm_and_si128(cmp, rc2);
// else
__m128i cmp3 = _mm_or_si128(cmp1, cmp2);
__m128i tmp4 = _mm_sub_epi32(da, dc);
tmp4 = Multiply32_SSE2(tmp4, sa);
tmp4 = shim_mm_div_epi32(tmp4, sc);
__m128i tmp5 = _mm_sub_epi32(da, SkMin32_SSE2(da, tmp4));
tmp5 = Multiply32_SSE2(sa, tmp5);
__m128i rc3 = _mm_add_epi32(tmp5, tmp2);
rc3 = _mm_add_epi32(rc3, tmp3);
rc3 = clamp_div255round_SSE2(rc3);
rc3 = _mm_andnot_si128(cmp3, rc3);
__m128i rc = _mm_or_si128(rc1, rc2);
rc = _mm_or_si128(rc, rc3);
return rc;
}
static inline __m128i colordodge_byte_SSE2(const __m128i& sc, const __m128i& dc,
const __m128i& sa, const __m128i& da) {
__m128i diff = _mm_sub_epi32(sa, sc);
__m128i ida = _mm_sub_epi32(_mm_set1_epi32(255), da);
__m128i isa = _mm_sub_epi32(_mm_set1_epi32(255), sa);
// if (0 == dc)
__m128i cmp1 = _mm_cmpeq_epi32(dc, _mm_setzero_si128());
__m128i rc1 = _mm_and_si128(cmp1, SkAlphaMulAlpha_SSE2(sc, ida));
// else if (0 == diff)
__m128i cmp2 = _mm_cmpeq_epi32(diff, _mm_setzero_si128());
__m128i cmp = _mm_andnot_si128(cmp1, cmp2);
__m128i tmp1 = _mm_mullo_epi16(sa, da);
__m128i tmp2 = _mm_mullo_epi16(sc, ida);
__m128i tmp3 = _mm_mullo_epi16(dc, isa);
__m128i rc2 = _mm_add_epi32(tmp1, tmp2);
rc2 = _mm_add_epi32(rc2, tmp3);
rc2 = clamp_div255round_SSE2(rc2);
rc2 = _mm_and_si128(cmp, rc2);
// else
__m128i cmp3 = _mm_or_si128(cmp1, cmp2);
__m128i value = _mm_mullo_epi16(dc, sa);
diff = shim_mm_div_epi32(value, diff);
__m128i tmp4 = SkMin32_SSE2(da, diff);
tmp4 = Multiply32_SSE2(sa, tmp4);
__m128i rc3 = _mm_add_epi32(tmp4, tmp2);
rc3 = _mm_add_epi32(rc3, tmp3);
rc3 = clamp_div255round_SSE2(rc3);
rc3 = _mm_andnot_si128(cmp3, rc3);
__m128i rc = _mm_or_si128(rc1, rc2);
rc = _mm_or_si128(rc, rc3);
return rc;
}
static inline __m128i blendfunc_multiply_byte_SSE2(const __m128i& sc, const __m128i& dc,
const __m128i& sa, const __m128i& da) {
// sc * (255 - da)
__m128i ret1 = _mm_sub_epi32(_mm_set1_epi32(255), da);
ret1 = _mm_mullo_epi16(sc, ret1);
// dc * (255 - sa)
__m128i ret2 = _mm_sub_epi32(_mm_set1_epi32(255), sa);
ret2 = _mm_mullo_epi16(dc, ret2);
// sc * dc
__m128i ret3 = _mm_mullo_epi16(sc, dc);
__m128i ret = _mm_add_epi32(ret1, ret2);
ret = _mm_add_epi32(ret, ret3);
return clamp_div255round_SSE2(ret);
}
// Portable version overlay_byte() is in SkXfermode.cpp.
static inline __m128i overlay_byte_SSE2(const __m128i& sc, const __m128i& dc,
const __m128i& sa, const __m128i& da) {
__m128i ida = _mm_sub_epi32(_mm_set1_epi32(255), da);
__m128i tmp1 = _mm_mullo_epi16(sc, ida);
__m128i isa = _mm_sub_epi32(_mm_set1_epi32(255), sa);
__m128i tmp2 = _mm_mullo_epi16(dc, isa);
__m128i tmp = _mm_add_epi32(tmp1, tmp2);
__m128i cmp = _mm_cmpgt_epi32(_mm_slli_epi32(dc, 1), da);
__m128i rc1 = _mm_slli_epi32(sc, 1); // 2 * sc
rc1 = Multiply32_SSE2(rc1, dc); // *dc
__m128i rc2 = _mm_mullo_epi16(sa, da); // sa * da
__m128i tmp3 = _mm_slli_epi32(_mm_sub_epi32(da, dc), 1); // 2 * (da - dc)
tmp3 = Multiply32_SSE2(tmp3, _mm_sub_epi32(sa, sc)); // * (sa - sc)
rc2 = _mm_sub_epi32(rc2, tmp3);
__m128i rc = _mm_or_si128(_mm_andnot_si128(cmp, rc1),
_mm_and_si128(cmp, rc2));
return clamp_div255round_SSE2(_mm_add_epi32(rc, tmp));
}
static inline __m128i softlight_byte_SSE2(const __m128i& sc, const __m128i& dc,
const __m128i& sa, const __m128i& da) {
__m128i tmp1, tmp2, tmp3;
// int m = da ? dc * 256 / da : 0;
__m128i cmp = _mm_cmpeq_epi32(da, _mm_setzero_si128());
__m128i m = _mm_slli_epi32(dc, 8);
__m128 x = _mm_cvtepi32_ps(m);
__m128 y = _mm_cvtepi32_ps(da);
m = _mm_cvttps_epi32(_mm_div_ps(x, y));
m = _mm_andnot_si128(cmp, m);
// if (2 * sc <= sa)
tmp1 = _mm_slli_epi32(sc, 1); // 2 * sc
__m128i cmp1 = _mm_cmpgt_epi32(tmp1, sa);
tmp1 = _mm_sub_epi32(tmp1, sa); // 2 * sc - sa
tmp2 = _mm_sub_epi32(_mm_set1_epi32(256), m); // 256 - m
tmp1 = Multiply32_SSE2(tmp1, tmp2);
tmp1 = _mm_srai_epi32(tmp1, 8);
tmp1 = _mm_add_epi32(sa, tmp1);
tmp1 = Multiply32_SSE2(dc, tmp1);
__m128i rc1 = _mm_andnot_si128(cmp1, tmp1);
// else if (4 * dc <= da)
tmp2 = _mm_slli_epi32(dc, 2); // dc * 4
__m128i cmp2 = _mm_cmpgt_epi32(tmp2, da);
__m128i i = _mm_slli_epi32(m, 2); // 4 * m
__m128i j = _mm_add_epi32(i, _mm_set1_epi32(256)); // 4 * m + 256
__m128i k = Multiply32_SSE2(i, j); // 4 * m * (4 * m + 256)
__m128i t = _mm_sub_epi32(m, _mm_set1_epi32(256)); // m - 256
i = Multiply32_SSE2(k, t); // 4 * m * (4 * m + 256) * (m - 256)
i = _mm_srai_epi32(i, 16); // >> 16
j = Multiply32_SSE2(_mm_set1_epi32(7), m); // 7 * m
tmp2 = _mm_add_epi32(i, j);
i = Multiply32_SSE2(dc, sa); // dc * sa
j = _mm_slli_epi32(sc, 1); // 2 * sc
j = _mm_sub_epi32(j, sa); // 2 * sc - sa
j = Multiply32_SSE2(da, j); // da * (2 * sc - sa)
tmp2 = Multiply32_SSE2(j, tmp2); // * tmp
tmp2 = _mm_srai_epi32(tmp2, 8); // >> 8
tmp2 = _mm_add_epi32(i, tmp2);
cmp = _mm_andnot_si128(cmp2, cmp1);
__m128i rc2 = _mm_and_si128(cmp, tmp2);
__m128i rc = _mm_or_si128(rc1, rc2);
// else
tmp3 = sqrt_unit_byte_SSE2(m);
tmp3 = _mm_sub_epi32(tmp3, m);
tmp3 = Multiply32_SSE2(j, tmp3); // j = da * (2 * sc - sa)
tmp3 = _mm_srai_epi32(tmp3, 8);
tmp3 = _mm_add_epi32(i, tmp3); // i = dc * sa
cmp = _mm_and_si128(cmp1, cmp2);
__m128i rc3 = _mm_and_si128(cmp, tmp3);
rc = _mm_or_si128(rc, rc3);
tmp1 = _mm_sub_epi32(_mm_set1_epi32(255), da); // 255 - da
tmp1 = _mm_mullo_epi16(sc, tmp1);
tmp2 = _mm_sub_epi32(_mm_set1_epi32(255), sa); // 255 - sa
tmp2 = _mm_mullo_epi16(dc, tmp2);
rc = _mm_add_epi32(rc, tmp1);
rc = _mm_add_epi32(rc, tmp2);
return clamp_div255round_SSE2(rc);
}