Lines Matching refs:summ

85 						__m128i summ, mull;  in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()  local
86 summ = _mm_madd_epi16(q11, _mm_loadu_si128((const __m128i*)(data+i-12))); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
87 …mm_madd_epi16(q10, _mm_loadu_si128((const __m128i*)(data+i-11))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
88 …_mm_madd_epi16(q9, _mm_loadu_si128((const __m128i*)(data+i-10))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
89 … _mm_madd_epi16(q8, _mm_loadu_si128((const __m128i*)(data+i-9))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
90 … _mm_madd_epi16(q7, _mm_loadu_si128((const __m128i*)(data+i-8))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
91 … _mm_madd_epi16(q6, _mm_loadu_si128((const __m128i*)(data+i-7))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
92 … _mm_madd_epi16(q5, _mm_loadu_si128((const __m128i*)(data+i-6))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
93 … _mm_madd_epi16(q4, _mm_loadu_si128((const __m128i*)(data+i-5))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
94 … _mm_madd_epi16(q3, _mm_loadu_si128((const __m128i*)(data+i-4))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
95 … _mm_madd_epi16(q2, _mm_loadu_si128((const __m128i*)(data+i-3))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
96 … _mm_madd_epi16(q1, _mm_loadu_si128((const __m128i*)(data+i-2))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
97 … _mm_madd_epi16(q0, _mm_loadu_si128((const __m128i*)(data+i-1))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
98 summ = _mm_sra_epi32(summ, cnt); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
99 …oreu_si128((__m128i*)(residual+i), _mm_sub_epi32(_mm_loadu_si128((const __m128i*)(data+i)), summ)); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
117 __m128i summ, mull; in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2() local
118 summ = _mm_madd_epi16(q10, _mm_loadu_si128((const __m128i*)(data+i-11))); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
119 …_mm_madd_epi16(q9, _mm_loadu_si128((const __m128i*)(data+i-10))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
120 … _mm_madd_epi16(q8, _mm_loadu_si128((const __m128i*)(data+i-9))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
121 … _mm_madd_epi16(q7, _mm_loadu_si128((const __m128i*)(data+i-8))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
122 … _mm_madd_epi16(q6, _mm_loadu_si128((const __m128i*)(data+i-7))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
123 … _mm_madd_epi16(q5, _mm_loadu_si128((const __m128i*)(data+i-6))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
124 … _mm_madd_epi16(q4, _mm_loadu_si128((const __m128i*)(data+i-5))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
125 … _mm_madd_epi16(q3, _mm_loadu_si128((const __m128i*)(data+i-4))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
126 … _mm_madd_epi16(q2, _mm_loadu_si128((const __m128i*)(data+i-3))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
127 … _mm_madd_epi16(q1, _mm_loadu_si128((const __m128i*)(data+i-2))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
128 … _mm_madd_epi16(q0, _mm_loadu_si128((const __m128i*)(data+i-1))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
129 summ = _mm_sra_epi32(summ, cnt); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
130 …oreu_si128((__m128i*)(residual+i), _mm_sub_epi32(_mm_loadu_si128((const __m128i*)(data+i)), summ)); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
149 __m128i summ, mull; in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2() local
150 summ = _mm_madd_epi16(q9, _mm_loadu_si128((const __m128i*)(data+i-10))); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
151 … _mm_madd_epi16(q8, _mm_loadu_si128((const __m128i*)(data+i-9))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
152 … _mm_madd_epi16(q7, _mm_loadu_si128((const __m128i*)(data+i-8))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
153 … _mm_madd_epi16(q6, _mm_loadu_si128((const __m128i*)(data+i-7))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
154 … _mm_madd_epi16(q5, _mm_loadu_si128((const __m128i*)(data+i-6))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
155 … _mm_madd_epi16(q4, _mm_loadu_si128((const __m128i*)(data+i-5))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
156 … _mm_madd_epi16(q3, _mm_loadu_si128((const __m128i*)(data+i-4))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
157 … _mm_madd_epi16(q2, _mm_loadu_si128((const __m128i*)(data+i-3))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
158 … _mm_madd_epi16(q1, _mm_loadu_si128((const __m128i*)(data+i-2))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
159 … _mm_madd_epi16(q0, _mm_loadu_si128((const __m128i*)(data+i-1))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
160 summ = _mm_sra_epi32(summ, cnt); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
161 …oreu_si128((__m128i*)(residual+i), _mm_sub_epi32(_mm_loadu_si128((const __m128i*)(data+i)), summ)); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
177 __m128i summ, mull; in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2() local
178 summ = _mm_madd_epi16(q8, _mm_loadu_si128((const __m128i*)(data+i-9))); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
179 … _mm_madd_epi16(q7, _mm_loadu_si128((const __m128i*)(data+i-8))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
180 … _mm_madd_epi16(q6, _mm_loadu_si128((const __m128i*)(data+i-7))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
181 … _mm_madd_epi16(q5, _mm_loadu_si128((const __m128i*)(data+i-6))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
182 … _mm_madd_epi16(q4, _mm_loadu_si128((const __m128i*)(data+i-5))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
183 … _mm_madd_epi16(q3, _mm_loadu_si128((const __m128i*)(data+i-4))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
184 … _mm_madd_epi16(q2, _mm_loadu_si128((const __m128i*)(data+i-3))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
185 … _mm_madd_epi16(q1, _mm_loadu_si128((const __m128i*)(data+i-2))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
186 … _mm_madd_epi16(q0, _mm_loadu_si128((const __m128i*)(data+i-1))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
187 summ = _mm_sra_epi32(summ, cnt); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
188 …oreu_si128((__m128i*)(residual+i), _mm_sub_epi32(_mm_loadu_si128((const __m128i*)(data+i)), summ)); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
207 __m128i summ, mull; in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2() local
208 summ = _mm_madd_epi16(q7, _mm_loadu_si128((const __m128i*)(data+i-8))); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
209 … _mm_madd_epi16(q6, _mm_loadu_si128((const __m128i*)(data+i-7))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
210 … _mm_madd_epi16(q5, _mm_loadu_si128((const __m128i*)(data+i-6))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
211 … _mm_madd_epi16(q4, _mm_loadu_si128((const __m128i*)(data+i-5))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
212 … _mm_madd_epi16(q3, _mm_loadu_si128((const __m128i*)(data+i-4))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
213 … _mm_madd_epi16(q2, _mm_loadu_si128((const __m128i*)(data+i-3))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
214 … _mm_madd_epi16(q1, _mm_loadu_si128((const __m128i*)(data+i-2))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
215 … _mm_madd_epi16(q0, _mm_loadu_si128((const __m128i*)(data+i-1))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
216 summ = _mm_sra_epi32(summ, cnt); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
217 …oreu_si128((__m128i*)(residual+i), _mm_sub_epi32(_mm_loadu_si128((const __m128i*)(data+i)), summ)); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
231 __m128i summ, mull; in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2() local
232 summ = _mm_madd_epi16(q6, _mm_loadu_si128((const __m128i*)(data+i-7))); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
233 … _mm_madd_epi16(q5, _mm_loadu_si128((const __m128i*)(data+i-6))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
234 … _mm_madd_epi16(q4, _mm_loadu_si128((const __m128i*)(data+i-5))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
235 … _mm_madd_epi16(q3, _mm_loadu_si128((const __m128i*)(data+i-4))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
236 … _mm_madd_epi16(q2, _mm_loadu_si128((const __m128i*)(data+i-3))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
237 … _mm_madd_epi16(q1, _mm_loadu_si128((const __m128i*)(data+i-2))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
238 … _mm_madd_epi16(q0, _mm_loadu_si128((const __m128i*)(data+i-1))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
239 summ = _mm_sra_epi32(summ, cnt); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
240 …oreu_si128((__m128i*)(residual+i), _mm_sub_epi32(_mm_loadu_si128((const __m128i*)(data+i)), summ)); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
255 __m128i summ, mull; in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2() local
256 summ = _mm_madd_epi16(q5, _mm_loadu_si128((const __m128i*)(data+i-6))); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
257 … _mm_madd_epi16(q4, _mm_loadu_si128((const __m128i*)(data+i-5))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
258 … _mm_madd_epi16(q3, _mm_loadu_si128((const __m128i*)(data+i-4))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
259 … _mm_madd_epi16(q2, _mm_loadu_si128((const __m128i*)(data+i-3))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
260 … _mm_madd_epi16(q1, _mm_loadu_si128((const __m128i*)(data+i-2))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
261 … _mm_madd_epi16(q0, _mm_loadu_si128((const __m128i*)(data+i-1))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
262 summ = _mm_sra_epi32(summ, cnt); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
263 …oreu_si128((__m128i*)(residual+i), _mm_sub_epi32(_mm_loadu_si128((const __m128i*)(data+i)), summ)); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
275 __m128i summ, mull; in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2() local
276 summ = _mm_madd_epi16(q4, _mm_loadu_si128((const __m128i*)(data+i-5))); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
277 … _mm_madd_epi16(q3, _mm_loadu_si128((const __m128i*)(data+i-4))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
278 … _mm_madd_epi16(q2, _mm_loadu_si128((const __m128i*)(data+i-3))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
279 … _mm_madd_epi16(q1, _mm_loadu_si128((const __m128i*)(data+i-2))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
280 … _mm_madd_epi16(q0, _mm_loadu_si128((const __m128i*)(data+i-1))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
281 summ = _mm_sra_epi32(summ, cnt); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
282 …oreu_si128((__m128i*)(residual+i), _mm_sub_epi32(_mm_loadu_si128((const __m128i*)(data+i)), summ)); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
297 __m128i summ, mull; in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2() local
298 summ = _mm_madd_epi16(q3, _mm_loadu_si128((const __m128i*)(data+i-4))); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
299 … _mm_madd_epi16(q2, _mm_loadu_si128((const __m128i*)(data+i-3))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
300 … _mm_madd_epi16(q1, _mm_loadu_si128((const __m128i*)(data+i-2))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
301 … _mm_madd_epi16(q0, _mm_loadu_si128((const __m128i*)(data+i-1))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
302 summ = _mm_sra_epi32(summ, cnt); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
303 …oreu_si128((__m128i*)(residual+i), _mm_sub_epi32(_mm_loadu_si128((const __m128i*)(data+i)), summ)); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
313 __m128i summ, mull; in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2() local
314 summ = _mm_madd_epi16(q2, _mm_loadu_si128((const __m128i*)(data+i-3))); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
315 … _mm_madd_epi16(q1, _mm_loadu_si128((const __m128i*)(data+i-2))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
316 … _mm_madd_epi16(q0, _mm_loadu_si128((const __m128i*)(data+i-1))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
317 summ = _mm_sra_epi32(summ, cnt); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
318 …oreu_si128((__m128i*)(residual+i), _mm_sub_epi32(_mm_loadu_si128((const __m128i*)(data+i)), summ)); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
329 __m128i summ, mull; in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2() local
330 summ = _mm_madd_epi16(q1, _mm_loadu_si128((const __m128i*)(data+i-2))); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
331 … _mm_madd_epi16(q0, _mm_loadu_si128((const __m128i*)(data+i-1))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
332 summ = _mm_sra_epi32(summ, cnt); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
333 …oreu_si128((__m128i*)(residual+i), _mm_sub_epi32(_mm_loadu_si128((const __m128i*)(data+i)), summ)); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
341 __m128i summ; in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2() local
342 summ = _mm_madd_epi16(q0, _mm_loadu_si128((const __m128i*)(data+i-1))); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
343 summ = _mm_sra_epi32(summ, cnt); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
344 …oreu_si128((__m128i*)(residual+i), _mm_sub_epi32(_mm_loadu_si128((const __m128i*)(data+i)), summ)); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()