39     __m128i sixteen = _mm_cvtsi32_si128(16);  in S32_opaque_D32_filter_DX_SSE2()
45     __m128i allY = _mm_cvtsi32_si128(subY);  in S32_opaque_D32_filter_DX_SSE2()
51     __m128i negY = _mm_sub_epi16(sixteen, allY);  in S32_opaque_D32_filter_DX_SSE2()
60     __m128i zero = _mm_setzero_si128();  in S32_opaque_D32_filter_DX_SSE2()
67         __m128i allX = _mm_cvtsi32_si128((XX >> 14) & 0x0F);  in S32_opaque_D32_filter_DX_SSE2()
76         __m128i negX = _mm_sub_epi16(sixteen, allX);  in S32_opaque_D32_filter_DX_SSE2()
79         __m128i a00 = _mm_cvtsi32_si128(row0[x0]);  in S32_opaque_D32_filter_DX_SSE2()
80         __m128i a01 = _mm_cvtsi32_si128(row0[x1]);  in S32_opaque_D32_filter_DX_SSE2()
81         __m128i a10 = _mm_cvtsi32_si128(row1[x0]);  in S32_opaque_D32_filter_DX_SSE2()
82         __m128i a11 = _mm_cvtsi32_si128(row1[x1]);  in S32_opaque_D32_filter_DX_SSE2()
85         __m128i a00a10 = _mm_unpacklo_epi32(a10, a00);  in S32_opaque_D32_filter_DX_SSE2()
97         __m128i a01a11 = _mm_unpacklo_epi32(a11, a01);  in S32_opaque_D32_filter_DX_SSE2()
109         __m128i sum = _mm_add_epi16(a00a10, a01a11);  in S32_opaque_D32_filter_DX_SSE2()
112         __m128i shifted = _mm_shuffle_epi32(sum, 0xEE);  in S32_opaque_D32_filter_DX_SSE2()
145     __m128i sixteen = _mm_cvtsi32_si128(16);  in S32_alpha_D32_filter_DX_SSE2()
151     __m128i allY = _mm_cvtsi32_si128(subY);  in S32_alpha_D32_filter_DX_SSE2()
157     __m128i negY = _mm_sub_epi16(sixteen, allY);  in S32_alpha_D32_filter_DX_SSE2()
166     __m128i zero = _mm_setzero_si128();  in S32_alpha_D32_filter_DX_SSE2()
169     __m128i alpha = _mm_set1_epi16(s.fAlphaScale);  in S32_alpha_D32_filter_DX_SSE2()
177         __m128i allX = _mm_cvtsi32_si128((XX >> 14) & 0x0F);  in S32_alpha_D32_filter_DX_SSE2()
186         __m128i negX = _mm_sub_epi16(sixteen, allX);  in S32_alpha_D32_filter_DX_SSE2()
189         __m128i a00 = _mm_cvtsi32_si128(row0[x0]);  in S32_alpha_D32_filter_DX_SSE2()
190         __m128i a01 = _mm_cvtsi32_si128(row0[x1]);  in S32_alpha_D32_filter_DX_SSE2()
191         __m128i a10 = _mm_cvtsi32_si128(row1[x0]);  in S32_alpha_D32_filter_DX_SSE2()
192         __m128i a11 = _mm_cvtsi32_si128(row1[x1]);  in S32_alpha_D32_filter_DX_SSE2()
195         __m128i a00a10 = _mm_unpacklo_epi32(a10, a00);  in S32_alpha_D32_filter_DX_SSE2()
207         __m128i a01a11 = _mm_unpacklo_epi32(a11, a01);  in S32_alpha_D32_filter_DX_SSE2()
219         __m128i sum = _mm_add_epi16(a00a10, a01a11);  in S32_alpha_D32_filter_DX_SSE2()
222         __m128i shifted = _mm_shuffle_epi32(sum, 0xEE);  in S32_alpha_D32_filter_DX_SSE2()