Lines Matching refs:__m128i
28 const __m128i zero = _mm_setzero_si128(); in DispatchAlpha()
29 const __m128i rgb_mask = _mm_set1_epi32(0xffffff00u); // to preserve RGB in DispatchAlpha()
30 const __m128i all_0xff = _mm_set_epi32(0, 0, ~0u, ~0u); in DispatchAlpha()
31 __m128i all_alphas = all_0xff; in DispatchAlpha()
39 __m128i* out = (__m128i*)dst; in DispatchAlpha()
42 const __m128i a0 = _mm_loadl_epi64((const __m128i*)&alpha[i]); in DispatchAlpha()
43 const __m128i a1 = _mm_unpacklo_epi8(a0, zero); in DispatchAlpha()
44 const __m128i a2_lo = _mm_unpacklo_epi16(a1, zero); in DispatchAlpha()
45 const __m128i a2_hi = _mm_unpackhi_epi16(a1, zero); in DispatchAlpha()
47 const __m128i b0_lo = _mm_loadu_si128(out + 0); in DispatchAlpha()
48 const __m128i b0_hi = _mm_loadu_si128(out + 1); in DispatchAlpha()
50 const __m128i b1_lo = _mm_and_si128(b0_lo, rgb_mask); in DispatchAlpha()
51 const __m128i b1_hi = _mm_and_si128(b0_hi, rgb_mask); in DispatchAlpha()
53 const __m128i b2_lo = _mm_or_si128(b1_lo, a2_lo); in DispatchAlpha()
54 const __m128i b2_hi = _mm_or_si128(b1_hi, a2_hi); in DispatchAlpha()
79 const __m128i zero = _mm_setzero_si128(); in DispatchAlphaToGreen()
83 const __m128i a0 = _mm_loadu_si128((const __m128i*)&alpha[i]); in DispatchAlphaToGreen()
84 const __m128i a1 = _mm_unpacklo_epi8(zero, a0); // note the 'zero' first! in DispatchAlphaToGreen()
85 const __m128i b1 = _mm_unpackhi_epi8(zero, a0); in DispatchAlphaToGreen()
86 const __m128i a2_lo = _mm_unpacklo_epi16(a1, zero); in DispatchAlphaToGreen()
87 const __m128i b2_lo = _mm_unpacklo_epi16(b1, zero); in DispatchAlphaToGreen()
88 const __m128i a2_hi = _mm_unpackhi_epi16(a1, zero); in DispatchAlphaToGreen()
89 const __m128i b2_hi = _mm_unpackhi_epi16(b1, zero); in DispatchAlphaToGreen()
90 _mm_storeu_si128((__m128i*)&dst[i + 0], a2_lo); in DispatchAlphaToGreen()
91 _mm_storeu_si128((__m128i*)&dst[i + 4], a2_hi); in DispatchAlphaToGreen()
92 _mm_storeu_si128((__m128i*)&dst[i + 8], b2_lo); in DispatchAlphaToGreen()
93 _mm_storeu_si128((__m128i*)&dst[i + 12], b2_hi); in DispatchAlphaToGreen()
108 const __m128i a_mask = _mm_set1_epi32(0xffu); // to preserve alpha in ExtractAlpha()
109 const __m128i all_0xff = _mm_set_epi32(0, 0, ~0u, ~0u); in ExtractAlpha()
110 __m128i all_alphas = all_0xff; in ExtractAlpha()
118 const __m128i* src = (const __m128i*)argb; in ExtractAlpha()
121 const __m128i a0 = _mm_loadu_si128(src + 0); in ExtractAlpha()
122 const __m128i a1 = _mm_loadu_si128(src + 1); in ExtractAlpha()
123 const __m128i b0 = _mm_and_si128(a0, a_mask); in ExtractAlpha()
124 const __m128i b1 = _mm_and_si128(a1, a_mask); in ExtractAlpha()
125 const __m128i c0 = _mm_packs_epi32(b0, b1); in ExtractAlpha()
126 const __m128i d0 = _mm_packus_epi16(c0, c0); in ExtractAlpha()
128 _mm_storel_epi64((__m128i*)&alpha[i], d0); in ExtractAlpha()
155 const __m128i argb0 = _mm_loadl_epi64((__m128i*)&(RGBX)); \
156 const __m128i argb1 = _mm_unpacklo_epi8(argb0, zero); \
157 const __m128i alpha0 = _mm_and_si128(argb1, MASK); \
158 const __m128i alpha1 = _mm_shufflelo_epi16(alpha0, SHUFFLE); \
159 const __m128i alpha2 = _mm_shufflehi_epi16(alpha1, SHUFFLE); \
161 const __m128i scale0 = _mm_mullo_epi16(alpha2, MULT); \
162 const __m128i scale1 = _mm_mulhi_epu16(alpha2, MULT); \
163 const __m128i argb2 = _mm_mulhi_epu16(argb1, scale0); \
164 const __m128i argb3 = _mm_mullo_epi16(argb1, scale1); \
165 const __m128i argb4 = _mm_adds_epu16(argb2, argb3); \
166 const __m128i argb5 = _mm_srli_epi16(argb4, 7); \
167 const __m128i argb6 = _mm_or_si128(argb5, alpha0); \
168 const __m128i argb7 = _mm_packus_epi16(argb6, zero); \
169 _mm_storel_epi64((__m128i*)&(RGBX), argb7); \
174 const __m128i zero = _mm_setzero_si128(); in ApplyAlphaMultiply()
181 const __m128i kMask = _mm_set_epi16(0xff, 0, 0, 0, 0xff, 0, 0, 0); in ApplyAlphaMultiply()
182 const __m128i kMult = in ApplyAlphaMultiply()
188 const __m128i kMask = _mm_set_epi16(0, 0, 0, 0xff, 0, 0, 0, 0xff); in ApplyAlphaMultiply()
189 const __m128i kMult = in ApplyAlphaMultiply()
225 const __m128i zero = _mm_setzero_si128(); in MultARGBRow()
226 const __m128i kRound = in MultARGBRow()
228 const __m128i kMult = in MultARGBRow()
230 const __m128i kOne64 = _mm_set_epi16(1u << 8, 0, 0, 0, 1u << 8, 0, 0, 0); in MultARGBRow()
233 const __m128i argb0 = _mm_loadl_epi64((__m128i*)&ptr[x]); in MultARGBRow()
234 const __m128i argb1 = _mm_unpacklo_epi8(argb0, zero); in MultARGBRow()
235 const __m128i tmp0 = _mm_shufflelo_epi16(argb1, _MM_SHUFFLE(3, 3, 3, 3)); in MultARGBRow()
236 const __m128i tmp1 = _mm_shufflehi_epi16(tmp0, _MM_SHUFFLE(3, 3, 3, 3)); in MultARGBRow()
237 const __m128i tmp2 = _mm_srli_epi64(tmp1, 16); in MultARGBRow()
238 const __m128i scale0 = _mm_mullo_epi16(tmp1, kMult); in MultARGBRow()
239 const __m128i scale1 = _mm_or_si128(tmp2, kOne64); in MultARGBRow()
240 const __m128i argb2 = _mm_mulhi_epu16(argb1, scale0); in MultARGBRow()
241 const __m128i argb3 = _mm_mullo_epi16(argb1, scale1); in MultARGBRow()
242 const __m128i argb4 = _mm_adds_epu16(argb2, argb3); in MultARGBRow()
243 const __m128i argb5 = _mm_adds_epu16(argb4, kRound); in MultARGBRow()
244 const __m128i argb6 = _mm_srli_epi16(argb5, 8); in MultARGBRow()
245 const __m128i argb7 = _mm_packus_epi16(argb6, zero); in MultARGBRow()
246 _mm_storel_epi64((__m128i*)&ptr[x], argb7); in MultARGBRow()
258 const __m128i zero = _mm_setzero_si128(); in MultRow()
259 const __m128i kRound = _mm_set1_epi16(1 << 7); in MultRow()
262 const __m128i v0 = _mm_loadl_epi64((__m128i*)&ptr[x]); in MultRow()
263 const __m128i v1 = _mm_unpacklo_epi8(v0, zero); in MultRow()
264 const __m128i alpha0 = _mm_loadl_epi64((const __m128i*)&alpha[x]); in MultRow()
265 const __m128i alpha1 = _mm_unpacklo_epi8(alpha0, zero); in MultRow()
266 const __m128i alpha2 = _mm_unpacklo_epi8(alpha0, alpha0); in MultRow()
267 const __m128i v2 = _mm_mulhi_epu16(v1, alpha2); in MultRow()
268 const __m128i v3 = _mm_mullo_epi16(v1, alpha1); in MultRow()
269 const __m128i v4 = _mm_adds_epu16(v2, v3); in MultRow()
270 const __m128i v5 = _mm_adds_epu16(v4, kRound); in MultRow()
271 const __m128i v6 = _mm_srli_epi16(v5, 8); in MultRow()
272 const __m128i v7 = _mm_packus_epi16(v6, zero); in MultRow()
273 _mm_storel_epi64((__m128i*)&ptr[x], v7); in MultRow()