/external/libhevc/common/x86/ |
ihevc_32x32_itrans_recon_sse42_intr.c | 196 __m128i m_coeff1, m_coeff2, m_coeff3, m_coeff4; local 344 m_coeff2 = _mm_loadu_si128((__m128i *)&g_ai2_ihevc_trans_32_intr_packed[4][0]); //87 387 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_10, m_coeff2); 401 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_14, m_coeff2); 600 m_coeff2 = _mm_loadu_si128((__m128i *)&g_ai2_ihevc_trans_16_even_packed[1][0]); //36 -83 655 m_coeff2 = _mm_loadu_si128((__m128i *)&g_ai2_ihevc_trans_16_even_packed[6][0]); //75 693 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_10, m_coeff2); 707 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_14, m_coeff2); 780 m_coeff2 = _mm_loadu_si128((__m128i *)&g_ai2_ihevc_trans_32_intr_packed[1][0]); //80 70 [all...] |
ihevc_itrans_recon_32x32_ssse3_intr.c | 204 __m128i m_coeff1, m_coeff2, m_coeff3, m_coeff4; local 353 m_coeff2 = _mm_load_si128((__m128i *)&g_ai2_ihevc_trans_32_intr_packed[4][0]); //87 398 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_10, m_coeff2); 412 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_14, m_coeff2); 611 m_coeff2 = _mm_load_si128((__m128i *)&g_ai2_ihevc_trans_16_even_packed[1][0]); //36 -83 666 m_coeff2 = _mm_load_si128((__m128i *)&g_ai2_ihevc_trans_16_even_packed[6][0]); //75 704 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_10, m_coeff2); 718 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_14, m_coeff2); 791 m_coeff2 = _mm_load_si128((__m128i *)&g_ai2_ihevc_trans_32_intr_packed[1][0]); //80 70 [all...] |
ihevc_itrans_recon_ssse3_intr.c | 232 //m_temp_reg_31 = _mm_mullo_epi32(m_temp_reg_11, m_coeff2);//55*c1 246 //m_temp_reg_33 = _mm_mullo_epi32(m_temp_reg_12, m_coeff2);//55*c2 254 //m_temp_reg_34 = _mm_mullo_epi32(m_temp_reg_10, m_coeff2);//55*c0 370 //m_temp_reg_31 = _mm_mullo_epi32(m_temp_reg_1, m_coeff2);//55*c1 384 //m_temp_reg_33 = _mm_mullo_epi32(m_temp_reg_2, m_coeff2);//55*c2 392 //m_temp_reg_34 = _mm_mullo_epi32(m_temp_reg_0, m_coeff2);//55*c0 1020 __m128i m_coeff1, m_coeff2, m_coeff3, m_coeff4; local [all...] |
ihevc_itrans_recon_sse42_intr.c | 144 __m128i m_coeff1, m_coeff2, m_coeff3; local 202 m_coeff2 = _mm_loadu_si128((__m128i *)&g_ai4_ihevc_trans_4_ttype1[0][0]); //55 207 m_temp_reg_31 = _mm_mullo_epi32(m_temp_reg_11, m_coeff2); //55*c1 212 m_temp_reg_33 = _mm_mullo_epi32(m_temp_reg_12, m_coeff2); //55*c2 216 m_temp_reg_34 = _mm_mullo_epi32(m_temp_reg_10, m_coeff2); //55*c0 298 m_temp_reg_31 = _mm_mullo_epi32(m_temp_reg_1, m_coeff2); //55*c1 303 m_temp_reg_33 = _mm_mullo_epi32(m_temp_reg_2, m_coeff2); //55*c2 307 m_temp_reg_34 = _mm_mullo_epi32(m_temp_reg_0, m_coeff2); //55*c0 845 __m128i m_coeff1, m_coeff2, m_coeff3, m_coeff4; local 882 m_coeff2 = _mm_loadu_si128((__m128i *)&g_ai2_ihevc_trans_intr_even_8[3][0]) [all...] |
ihevc_16x16_itrans_recon_sse42_intr.c | 160 __m128i m_coeff1, m_coeff2, m_coeff3, m_coeff4; local 401 m_coeff2 = _mm_loadu_si128((__m128i *)&g_ai2_ihevc_trans_16_even[7][0]); //36 -83 407 m_temp_reg_22 = _mm_madd_epi16(m_temp_reg_0, m_coeff2); 410 m_temp_reg_23 = _mm_madd_epi16(m_temp_reg_1, m_coeff2); 434 m_coeff2 = _mm_loadu_si128((__m128i *)&g_ai2_ihevc_trans_16_even[4][0]); //50 18 522 m_coeff2 = _mm_loadu_si128((__m128i *)&g_ai2_ihevc_trans_16_even[10][0]); //18 75 599 m_coeff2 = _mm_loadu_si128((__m128i *)&g_ai2_ihevc_trans_16_even[7][0]); //36 -83 605 m_temp_reg_22 = _mm_madd_epi16(m_temp_reg_0, m_coeff2); 608 m_temp_reg_23 = _mm_madd_epi16(m_temp_reg_1, m_coeff2); 629 m_coeff2 = _mm_loadu_si128((__m128i *)&g_ai2_ihevc_trans_16_even[4][0]); //50 1 [all...] |
ihevc_itrans_recon_16x16_ssse3_intr.c | 164 __m128i m_coeff1, m_coeff2, m_coeff3, m_coeff4; local 407 m_coeff2 = _mm_load_si128((__m128i *)&g_ai2_ihevc_trans_16_even[7][0]); //36 -83 413 m_temp_reg_22 = _mm_madd_epi16(m_temp_reg_0, m_coeff2); 416 m_temp_reg_23 = _mm_madd_epi16(m_temp_reg_1, m_coeff2); 438 m_coeff2 = _mm_load_si128((__m128i *)&g_ai2_ihevc_trans_16_even[4][0]); //50 18 526 m_coeff2 = _mm_load_si128((__m128i *)&g_ai2_ihevc_trans_16_even[10][0]); //18 75 603 m_coeff2 = _mm_load_si128((__m128i *)&g_ai2_ihevc_trans_16_even[7][0]); //36 -83 609 m_temp_reg_22 = _mm_madd_epi16(m_temp_reg_0, m_coeff2); 612 m_temp_reg_23 = _mm_madd_epi16(m_temp_reg_1, m_coeff2); 633 m_coeff2 = _mm_load_si128((__m128i *)&g_ai2_ihevc_trans_16_even[4][0]); //50 1 [all...] |
/external/libmpeg2/common/x86/ |
impeg2_idct_recon_sse42_intr.c | 183 __m128i m_coeff1, m_coeff2, m_coeff3, m_coeff4; local 220 m_coeff2 = _mm_loadu_si128((__m128i *)&gai2_impeg2_idct_even_8_q15[3][0]); 226 m_temp_reg_12 = _mm_madd_epi16(m_temp_reg_0, m_coeff2); 237 m_coeff2 = _mm_loadu_si128((__m128i *)&gai2_impeg2_idct_even_8_q15[2][0]); //add 2B*83+6B*36 ,2T*83+6T*36 244 m_temp_reg_14 = _mm_madd_epi16(m_temp_reg_4, m_coeff2); 253 m_coeff2 = _mm_loadu_si128((__m128i *)&gai2_impeg2_idct_odd_8_q15[1][0]); 313 m_coeff2 = _mm_loadu_si128((__m128i *)&gai2_impeg2_idct_odd_8_q15[5][0]); 428 m_coeff2 = _mm_loadu_si128((__m128i *)&gai2_impeg2_idct_even_8_q15[3][0]); 434 m_temp_reg_12 = _mm_madd_epi16(m_temp_reg_0, m_coeff2); 445 m_coeff2 = _mm_loadu_si128((__m128i *)&gai2_impeg2_idct_even_8_q15[2][0]); //add 2B*83+6B*36 ,2T*83+6T*3 [all...] |