/****************************************************************************** * * Copyright (C) 2012 Ittiam Systems Pvt Ltd, Bangalore * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at: * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. * ******************************************************************************/ /** ******************************************************************************* * @file * ihevc_deblck_atom_intr.c * * @brief * Contains function definitions for deblocking filters * * @author * Rishab * * @par List of Functions: * - ihevc_deblk_luma_vert_ssse3() * - ihevc_deblk_luma_horz_ssse3() * - ihevc_deblk_chroma_vert_ssse3() * - ihevc_deblk_chroma_horz_ssse3() * * @remarks * None * ******************************************************************************* */ #include #include #include #include "ihevc_typedefs.h" #include "ihevc_platform_macros.h" #include "ihevc_macros.h" #include "ihevc_deblk.h" #include "ihevc_deblk_tables.h" #include "ihevc_debug.h" #include "ihevc_tables_x86_intr.h" #include /** ******************************************************************************* * * @brief * Decision process and filtering for the luma block vertical edge. * * @par Description: * The decision process for the luma block vertical edge is carried out and * an appropriate filter is applied. The boundary filter strength, bs should * be greater than 0. The pcm flags and the transquant bypass flags should * be taken care of by the calling function. * * @param[in] pu1_src * Pointer to the src sample q(0,0) * * @param[in] src_strd * Source stride * * @param[in] bs * Boundary filter strength of q(0,0) * * @param[in] quant_param_p * quantization parameter of p block * * @param[in] quant_param_q * quantization parameter of p block * * @param[in] beta_offset_div2 * * * @param[in] tc_offset_div2 * * * @param[in] filter_flag_p * flag whether to filter the p block * * @param[in] filter_flag_q * flag whether to filter the q block * * @returns * * @remarks * None * ******************************************************************************* */ void ihevc_deblk_luma_vert_ssse3(UWORD8 *pu1_src, WORD32 src_strd, WORD32 bs, WORD32 quant_param_p, WORD32 quant_param_q, WORD32 beta_offset_div2, WORD32 tc_offset_div2, WORD32 filter_flag_p, WORD32 filter_flag_q) { WORD32 qp_luma, beta_indx, tc_indx; WORD32 beta, tc; WORD32 d, dp, dq, d_sam0, d_sam3; WORD32 d3, d0, de_0, de_1, de_2, de_3; WORD32 de, dep, deq; __m128i src_row0_8x16b, src_row1_8x16b, src_row2_8x16b, src_row3_8x16b; { __m128i src_tmp_8x16b, coef_8x16b, mask_d_result_4x32b, mask_de_result_8x16b; __m128i mask_16x8b, temp_coef0_8x16b, temp_coef1_8x16b; ASSERT((bs > 0) && (bs <= 3)); ASSERT(filter_flag_p || filter_flag_q); qp_luma = (quant_param_p + quant_param_q + 1) >> 1; beta_indx = CLIP3(qp_luma + (beta_offset_div2 << 1), 0, 51); /* BS based on implementation can take value 3 if it is intra/inter egde */ /* based on BS, tc index is calcuated by adding 2 * ( bs - 1) to QP and tc_offset */ /* for BS = 1 adding factor is (0*2), BS = 2 or 3 adding factor is (1*2) */ /* the above desired functionallity is achieved by doing (2*(bs>>1)) */ tc_indx = CLIP3(qp_luma + (2 * (bs >> 1)) + (tc_offset_div2 << 1), 0, 53); beta = gai4_ihevc_beta_table[beta_indx]; tc = gai4_ihevc_tc_table[tc_indx]; if(0 == tc) { return; } src_row0_8x16b = _mm_loadl_epi64((__m128i *)(pu1_src - 4)); src_row3_8x16b = _mm_loadl_epi64((__m128i *)((pu1_src - 4) + 3 * src_strd)); coef_8x16b = _mm_load_si128((__m128i *)(coef_d)); mask_16x8b = _mm_load_si128((__m128i *)(shuffle_d)); src_tmp_8x16b = _mm_unpacklo_epi64(src_row0_8x16b, src_row3_8x16b); mask_de_result_8x16b = _mm_shuffle_epi8(src_tmp_8x16b, mask_16x8b); mask_d_result_4x32b = _mm_maddubs_epi16(src_tmp_8x16b, coef_8x16b); //to get all 1's of 8 bit in (1) temp_coef0_8x16b = _mm_cmpeq_epi16(src_tmp_8x16b, src_tmp_8x16b); temp_coef1_8x16b = _mm_srli_epi16(temp_coef0_8x16b, 15); //accumulating values foe dp3 dq3 , dp0 dq0 values mask_d_result_4x32b = _mm_madd_epi16(mask_d_result_4x32b, temp_coef1_8x16b); temp_coef1_8x16b = _mm_packus_epi16(temp_coef1_8x16b, temp_coef1_8x16b); // to get all 1,-1 sets of 16 bits in (0) temp_coef0_8x16b = _mm_unpacklo_epi8(temp_coef0_8x16b, temp_coef1_8x16b); //q33-q30,p33-p30,q03-q00,p03-p00,0,q30-p30,0,q00-p00 mask_de_result_8x16b = _mm_maddubs_epi16(mask_de_result_8x16b, temp_coef0_8x16b); //to get 16 bit 1's temp_coef0_8x16b = _mm_srli_epi16(temp_coef1_8x16b, 8); // dq3 dp3 dq0 dp0 mask_d_result_4x32b = _mm_abs_epi32(mask_d_result_4x32b); mask_16x8b = _mm_shuffle_epi32(mask_d_result_4x32b, 0xec); mask_d_result_4x32b = _mm_shuffle_epi32(mask_d_result_4x32b, 0x49); // dq dp d3 d0 mask_d_result_4x32b = _mm_add_epi32(mask_d_result_4x32b, mask_16x8b); //|q33-q30|,|p33-p30|,|q03-q00|,|p03-p00|,0,|q30-p30|,0,|q00-p00| mask_de_result_8x16b = _mm_abs_epi16(mask_de_result_8x16b); //|q33-q30|+|p33-p30|,|q03-q00|+|p03-p00|,0+|q30-p30|,0+|q00-p00| mask_de_result_8x16b = _mm_madd_epi16(mask_de_result_8x16b, temp_coef0_8x16b); ///store back in a single variable temp_coef0_8x16b = _mm_srli_si128(mask_d_result_4x32b, 4); temp_coef1_8x16b = _mm_srli_si128(mask_d_result_4x32b, 8); mask_16x8b = _mm_srli_si128(mask_d_result_4x32b, 12); d0 = _mm_cvtsi128_si32(mask_d_result_4x32b); d3 = _mm_cvtsi128_si32(temp_coef0_8x16b); dp = _mm_cvtsi128_si32(temp_coef1_8x16b); dq = _mm_cvtsi128_si32(mask_16x8b); //getting d d = d0 + d3; ///store back in a single variable temp_coef0_8x16b = _mm_srli_si128(mask_de_result_8x16b, 4); temp_coef1_8x16b = _mm_srli_si128(mask_de_result_8x16b, 8); mask_16x8b = _mm_srli_si128(mask_de_result_8x16b, 12); de_0 = _mm_cvtsi128_si32(mask_de_result_8x16b); de_1 = _mm_cvtsi128_si32(temp_coef0_8x16b); de_2 = _mm_cvtsi128_si32(temp_coef1_8x16b); de_3 = _mm_cvtsi128_si32(mask_16x8b); de = 0; dep = 0; deq = 0; if(d < beta) { d_sam0 = 0; if((2 * d0 < (beta >> 2)) && (de_2 < (beta >> 3)) && (de_0 < ((5 * tc + 1) >> 1))) { d_sam0 = 1; } d_sam3 = 0; if((2 * d3 < (beta >> 2)) && (de_3 < (beta >> 3)) && de_1 < ((5 * tc + 1) >> 1)) { d_sam3 = 1; } de = (d_sam0 & d_sam3) + 1; dep = (dp < (beta + (beta >> 1)) >> 3) ? 1 : 0; deq = (dq < (beta + (beta >> 1)) >> 3) ? 1 : 0; if(tc <= 1) { dep = 0; deq = 0; } } } if(de != 0) { src_row1_8x16b = _mm_loadl_epi64((__m128i *)((pu1_src - 4) + src_strd)); src_row2_8x16b = _mm_loadl_epi64((__m128i *)((pu1_src - 4) + 2 * src_strd)); if(de == 2) { __m128i temp_pq_str0_16x8b; __m128i temp_pq1_str0_16x8b, temp_pq1_str1_16x8b; __m128i temp_pq2_str0_16x8b; __m128i temp_pq_str1_16x8b; __m128i temp_str0_16x8b, temp_str1_16x8b, temp_str2_16x8b, temp_str3_16x8b; __m128i temp_max0_16x8b, temp_max1_16x8b, temp_min0_16x8b, temp_min1_16x8b; __m128i const2_8x16b, const2tc_8x16b; LWORD64 mask, tc2; tc = tc << 1; mask = (((LWORD64)filter_flag_q) << 63) | (((LWORD64)filter_flag_p) << 31); tc2 = ((LWORD64)tc); const2_8x16b = _mm_cmpeq_epi16(src_row0_8x16b, src_row0_8x16b); //q'0-q'1-2 ,p'0-p'1-2 src_row0_8x16b = _mm_unpacklo_epi64(src_row0_8x16b, src_row2_8x16b); src_row1_8x16b = _mm_unpacklo_epi64(src_row1_8x16b, src_row3_8x16b); const2_8x16b = _mm_srli_epi16(const2_8x16b, 15); temp_pq_str0_16x8b = _mm_srli_epi64(src_row0_8x16b, 16); temp_pq_str1_16x8b = _mm_srli_epi64(src_row1_8x16b, 16); //arranged x x x x x x x x q31 q30 q1 q10 p30 p31 p10 p11 , x x x x x x x x q21 q20 q01 q00 p20 p21 p00 p01 temp_str0_16x8b = _mm_unpacklo_epi16(temp_pq_str0_16x8b, temp_pq_str1_16x8b); temp_str1_16x8b = _mm_unpackhi_epi16(temp_pq_str0_16x8b, temp_pq_str1_16x8b); const2_8x16b = _mm_packus_epi16(const2_8x16b, const2_8x16b); //arranged q31 q30 q21 q20 q1 q10 q01 q00 p30 p31 p20 p21 p10 p11 p00 p01 temp_pq_str0_16x8b = _mm_unpacklo_epi32(temp_str0_16x8b, temp_str1_16x8b); temp_pq_str0_16x8b = _mm_maddubs_epi16(temp_pq_str0_16x8b, const2_8x16b); //q'1-2, p'1-2 temp_pq1_str0_16x8b = _mm_srli_epi64(src_row0_8x16b, 8); temp_pq1_str1_16x8b = _mm_srli_epi64(src_row1_8x16b, 8); temp_str2_16x8b = _mm_unpacklo_epi16(temp_pq1_str0_16x8b, temp_pq1_str1_16x8b); temp_str3_16x8b = _mm_unpackhi_epi16(temp_pq1_str0_16x8b, temp_pq1_str1_16x8b); temp_str2_16x8b = _mm_shuffle_epi32(temp_str2_16x8b, 0x58); temp_str3_16x8b = _mm_shuffle_epi32(temp_str3_16x8b, 0x58); // q30 p30 q20 p20 q10 p10 q01 q00 p30 q20 p20 q10 p10 q01 q00 p00 temp_pq1_str0_16x8b = _mm_unpackhi_epi32(temp_str2_16x8b, temp_str3_16x8b); // q32 q31 q22 q21 q12 q11 q02 q01 p32 p31 p22 p21 p12 p11 p02 p01 temp_pq1_str1_16x8b = _mm_unpacklo_epi32(temp_str2_16x8b, temp_str3_16x8b); temp_pq1_str0_16x8b = _mm_maddubs_epi16(temp_pq1_str0_16x8b, const2_8x16b); temp_pq1_str1_16x8b = _mm_maddubs_epi16(temp_pq1_str1_16x8b, const2_8x16b); //clipping mask design temp_str1_16x8b = _mm_setzero_si128(); temp_str0_16x8b = _mm_loadl_epi64((__m128i *)(&mask)); const2tc_8x16b = _mm_loadl_epi64((__m128i *)(&tc2)); temp_str0_16x8b = _mm_shuffle_epi32(temp_str0_16x8b, 0x44); const2tc_8x16b = _mm_shuffle_epi8(const2tc_8x16b, temp_str1_16x8b); //clipping mask design temp_str0_16x8b = _mm_srai_epi32(temp_str0_16x8b, 31); const2tc_8x16b = _mm_and_si128(const2tc_8x16b, temp_str0_16x8b); //calculating Clipping MAX for all pixel values. temp_max0_16x8b = _mm_adds_epu8(src_row0_8x16b, const2tc_8x16b); temp_max1_16x8b = _mm_adds_epu8(src_row1_8x16b, const2tc_8x16b); //q'2-q'0-2,p'2-p'0-2 temp_pq2_str0_16x8b = _mm_unpacklo_epi16(src_row0_8x16b, src_row2_8x16b); temp_str3_16x8b = _mm_unpacklo_epi16(src_row1_8x16b, src_row3_8x16b); temp_pq2_str0_16x8b = _mm_shuffle_epi32(temp_pq2_str0_16x8b, 0x5c); temp_str3_16x8b = _mm_shuffle_epi32(temp_str3_16x8b, 0x5c); const2_8x16b = _mm_slli_epi16(const2_8x16b, 1); //arranged q33 q32 q23 q22 q13 q12 q03 q02 p33 p32 p23 p22 p13 p12 p03 p02 temp_str3_16x8b = _mm_unpacklo_epi16(temp_pq2_str0_16x8b, temp_str3_16x8b); temp_pq2_str0_16x8b = _mm_maddubs_epi16(temp_str3_16x8b, const2_8x16b); //calculating Clipping MIN for all pixel values. temp_min0_16x8b = _mm_subs_epu8(src_row0_8x16b, const2tc_8x16b); temp_min1_16x8b = _mm_subs_epu8(src_row1_8x16b, const2tc_8x16b); //q'0-q'1-2 ,p'0-p'1-2 temp_pq_str1_16x8b = _mm_shuffle_epi32(temp_pq_str0_16x8b, 0x4e); temp_pq_str0_16x8b = _mm_add_epi16(temp_pq_str0_16x8b, temp_pq_str1_16x8b); //q'1-2 p'1-2 temp_pq1_str0_16x8b = _mm_add_epi16(temp_pq1_str0_16x8b, temp_pq1_str1_16x8b); //to get 2 in 16 bit const2_8x16b = _mm_srli_epi16(const2_8x16b, 8); //to get q33 q23 q13 q03, p33 p23 p13 p03 temp_pq1_str1_16x8b = _mm_slli_epi16(temp_str3_16x8b, 8); temp_pq_str1_16x8b = _mm_srli_epi16(temp_str3_16x8b, 8); temp_pq1_str1_16x8b = _mm_srli_epi16(temp_pq1_str1_16x8b, 8); //q'1, p'1 (adding 2) temp_pq1_str0_16x8b = _mm_add_epi16(temp_pq1_str0_16x8b, const2_8x16b); //q'0-q'1,p'0-p'1 temp_pq_str0_16x8b = _mm_add_epi16(temp_pq_str0_16x8b, const2_8x16b); //q'2-q'1,p'2-p'1 temp_pq2_str0_16x8b = _mm_add_epi16(temp_pq2_str0_16x8b, const2_8x16b); //q'0 = (q'0-q'1)+q'1 ,p'0 = (p'0-p'1)+p'1; temp_pq_str0_16x8b = _mm_add_epi16(temp_pq1_str0_16x8b, temp_pq_str0_16x8b); //q'2 = (q'2-q'1)+q'1 ,p'2 = (p'2-p'1)+p'1; temp_pq2_str0_16x8b = _mm_add_epi16(temp_pq1_str0_16x8b, temp_pq2_str0_16x8b); //normalisation of all modified pixels temp_pq_str0_16x8b = _mm_srai_epi16(temp_pq_str0_16x8b, 3); temp_pq1_str0_16x8b = _mm_srai_epi16(temp_pq1_str0_16x8b, 2); temp_pq2_str0_16x8b = _mm_srai_epi16(temp_pq2_str0_16x8b, 3); //getting p0 p1 together and p2 p3 together temp_str0_16x8b = _mm_unpacklo_epi16(temp_pq1_str0_16x8b, temp_pq_str0_16x8b); temp_str2_16x8b = _mm_unpacklo_epi16(temp_pq1_str1_16x8b, temp_pq2_str0_16x8b); //getting q1 q0 together and q3 q2 together temp_pq_str0_16x8b = _mm_unpackhi_epi16(temp_pq_str0_16x8b, temp_pq1_str0_16x8b); temp_pq2_str0_16x8b = _mm_unpackhi_epi16(temp_pq2_str0_16x8b, temp_pq_str1_16x8b); //getting p's of row0 row1 together and of row2 row3 together temp_pq_str1_16x8b = _mm_unpacklo_epi32(temp_str2_16x8b, temp_str0_16x8b); temp_str2_16x8b = _mm_unpackhi_epi32(temp_str2_16x8b, temp_str0_16x8b); //getting q's of row0 row1 together and of row2 row3 together temp_str0_16x8b = _mm_unpacklo_epi32(temp_pq_str0_16x8b, temp_pq2_str0_16x8b); temp_pq_str0_16x8b = _mm_unpackhi_epi32(temp_pq_str0_16x8b, temp_pq2_str0_16x8b); //getting values for respective rows in 16 bit src_row0_8x16b = _mm_unpacklo_epi64(temp_pq_str1_16x8b, temp_str0_16x8b); src_row1_8x16b = _mm_unpackhi_epi64(temp_pq_str1_16x8b, temp_str0_16x8b); src_row2_8x16b = _mm_unpacklo_epi64(temp_str2_16x8b, temp_pq_str0_16x8b); src_row3_8x16b = _mm_unpackhi_epi64(temp_str2_16x8b, temp_pq_str0_16x8b); //packing values to 8 bit src_row0_8x16b = _mm_packus_epi16(src_row0_8x16b, src_row2_8x16b); src_row1_8x16b = _mm_packus_epi16(src_row1_8x16b, src_row3_8x16b); //Clipping MAX src_row0_8x16b = _mm_min_epu8(src_row0_8x16b, temp_max0_16x8b); src_row1_8x16b = _mm_min_epu8(src_row1_8x16b, temp_max1_16x8b); //Clipping MIN src_row0_8x16b = _mm_max_epu8(src_row0_8x16b, temp_min0_16x8b); src_row1_8x16b = _mm_max_epu8(src_row1_8x16b, temp_min1_16x8b); //separating row 2 and row 3 src_row2_8x16b = _mm_srli_si128(src_row0_8x16b, 8); src_row3_8x16b = _mm_srli_si128(src_row1_8x16b, 8); } else { __m128i tmp_delta0_8x16b, tmp_delta1_8x16b, tmp_delta2_8x16b, tmp_delta3_8x16b; __m128i tmp0_const_8x16b, tmp1_const_8x16b, tmp2_const_8x16b, tmp3_const_8x16b; __m128i coefdelta_0_8x16b, mask_pq_8x16b; __m128i const2_8x16b, consttc_8x16b; LWORD64 mask1; mask1 = (((LWORD64)(filter_flag_q & deq)) << 63) | (((LWORD64)filter_flag_q) << 47) | (((LWORD64)filter_flag_p) << 31) | (((LWORD64)(filter_flag_p & dep)) << 15); consttc_8x16b = _mm_set1_epi32(tc); src_row0_8x16b = _mm_unpacklo_epi64(src_row0_8x16b, src_row1_8x16b); src_row2_8x16b = _mm_unpacklo_epi64(src_row2_8x16b, src_row3_8x16b); tmp_delta2_8x16b = _mm_srli_epi64(src_row0_8x16b, 16); tmp_delta3_8x16b = _mm_srli_epi64(src_row2_8x16b, 16); tmp_delta2_8x16b = _mm_shuffle_epi32(tmp_delta2_8x16b, 0x08); tmp_delta3_8x16b = _mm_shuffle_epi32(tmp_delta3_8x16b, 0x08); //arranged q31 q30 p30 p31 q21 q20 p20 p21 q1 q10 p10 p11 q01 q00 p00 p01 tmp_delta2_8x16b = _mm_unpacklo_epi64(tmp_delta2_8x16b, tmp_delta3_8x16b); coefdelta_0_8x16b = _mm_load_si128((__m128i *)coef_de1); // (-3q1+9q0),(-9p0+3p1) tmp_delta3_8x16b = _mm_maddubs_epi16(tmp_delta2_8x16b, coefdelta_0_8x16b); //converting to 16 bit consttc_8x16b = _mm_packs_epi32(consttc_8x16b, consttc_8x16b); //getting -tc store tmp1_const_8x16b = _mm_cmpeq_epi32(consttc_8x16b, consttc_8x16b); //calc 10 *tc = 2*tc +8*tc ; 2*tc tmp2_const_8x16b = _mm_slli_epi16(consttc_8x16b, 1); //calc 10 *tc = 2*tc +8*tc ; 8*tc tmp0_const_8x16b = _mm_slli_epi16(consttc_8x16b, 3); //getting -tc store tmp3_const_8x16b = _mm_sign_epi16(consttc_8x16b, tmp1_const_8x16b); //calc 10 *tc tmp2_const_8x16b = _mm_add_epi16(tmp2_const_8x16b, tmp0_const_8x16b); //const 1 const2_8x16b = _mm_srli_epi16(tmp1_const_8x16b, 15); tmp_delta0_8x16b = _mm_madd_epi16(tmp_delta3_8x16b, const2_8x16b); const2_8x16b = _mm_srli_epi32(tmp1_const_8x16b, 31); //getting the mask values mask_pq_8x16b = _mm_loadl_epi64((__m128i *)(&mask1)); //loaded coef for delta1 calculation coefdelta_0_8x16b = _mm_load_si128((__m128i *)coef_dep1); //(-2q1+q0),(p0-2p1) tmp_delta3_8x16b = _mm_maddubs_epi16(tmp_delta2_8x16b, coefdelta_0_8x16b); //const 8 const2_8x16b = _mm_slli_epi32(const2_8x16b, 3); //rearranging the mask values mask_pq_8x16b = _mm_unpacklo_epi64(mask_pq_8x16b, mask_pq_8x16b); //normalisation of the filter tmp_delta0_8x16b = _mm_add_epi32(tmp_delta0_8x16b, const2_8x16b); tmp_delta0_8x16b = _mm_srai_epi32(tmp_delta0_8x16b, 4); //getting deltaq0 tmp_delta2_8x16b = _mm_sign_epi32(tmp_delta0_8x16b, tmp1_const_8x16b); //packing d3q d2q d1q d0q d3p d2p d1p d0p tmp_delta0_8x16b = _mm_packs_epi32(tmp_delta0_8x16b, tmp_delta2_8x16b); //absolute delta tmp_delta2_8x16b = _mm_abs_epi16(tmp_delta0_8x16b); //Clipping of delta0 tmp_delta0_8x16b = _mm_min_epi16(tmp_delta0_8x16b, consttc_8x16b); //mask for |delta| < 10*tc tmp0_const_8x16b = _mm_cmpgt_epi16(tmp2_const_8x16b, tmp_delta2_8x16b); //Clipping of delta0 tmp_delta0_8x16b = _mm_max_epi16(tmp_delta0_8x16b, tmp3_const_8x16b); //delta 1 calc starts //getting q32 q22 q12 q02 p32 p12 p22 p02 tmp2_const_8x16b = _mm_loadl_epi64((__m128i *)(shuffle0)); tmp_delta2_8x16b = _mm_shuffle_epi8(src_row0_8x16b, tmp2_const_8x16b); tmp_delta1_8x16b = _mm_shuffle_epi8(src_row2_8x16b, tmp2_const_8x16b); tmp_delta1_8x16b = _mm_unpacklo_epi32(tmp_delta2_8x16b, tmp_delta1_8x16b); //constant 1 const2_8x16b = _mm_srli_epi16(tmp1_const_8x16b, 15); //tc>>1 16 bit consttc_8x16b = _mm_srai_epi16(consttc_8x16b, 1); //getting -tc>>1 store 16 bit tmp1_const_8x16b = _mm_sign_epi16(consttc_8x16b, tmp1_const_8x16b); //2*delta0 tmp2_const_8x16b = _mm_add_epi16(tmp_delta0_8x16b, tmp_delta0_8x16b); //getting all respective q's and p's together tmp3_const_8x16b = _mm_load_si128((__m128i *)(shuffle1)); tmp_delta3_8x16b = _mm_shuffle_epi8(tmp_delta3_8x16b, tmp3_const_8x16b); //final adds for deltap1 and deltaq1 tmp_delta3_8x16b = _mm_add_epi16(tmp_delta3_8x16b, const2_8x16b); tmp_delta1_8x16b = _mm_add_epi16(tmp_delta1_8x16b, tmp2_const_8x16b); tmp_delta1_8x16b = _mm_add_epi16(tmp_delta1_8x16b, tmp_delta3_8x16b); tmp2_const_8x16b = _mm_setzero_si128(); tmp_delta1_8x16b = _mm_srai_epi16(tmp_delta1_8x16b, 2); // clipping delta1 tmp_delta1_8x16b = _mm_min_epi16(tmp_delta1_8x16b, consttc_8x16b); // clipping delta1 tmp_delta1_8x16b = _mm_max_epi16(tmp_delta1_8x16b, tmp1_const_8x16b); //getting the mask ready mask_pq_8x16b = _mm_srai_epi16(mask_pq_8x16b, 15); //masking of the delta values |delta|<10*tc tmp_delta1_8x16b = _mm_and_si128(tmp_delta1_8x16b, tmp0_const_8x16b); tmp_delta0_8x16b = _mm_and_si128(tmp_delta0_8x16b, tmp0_const_8x16b); //packing dq1 dq0 dp0 dp1 tmp1_const_8x16b = _mm_unpacklo_epi16(tmp_delta1_8x16b, tmp_delta0_8x16b); tmp_delta0_8x16b = _mm_unpackhi_epi16(tmp_delta0_8x16b, tmp_delta1_8x16b); tmp_delta1_8x16b = _mm_unpackhi_epi32(tmp1_const_8x16b, tmp_delta0_8x16b); tmp_delta0_8x16b = _mm_unpacklo_epi32(tmp1_const_8x16b, tmp_delta0_8x16b); //masking of the delta values dep, deq , filter_p ,filter_q tmp_delta0_8x16b = _mm_and_si128(tmp_delta0_8x16b, mask_pq_8x16b); tmp_delta1_8x16b = _mm_and_si128(tmp_delta1_8x16b, mask_pq_8x16b); //converting 8bit to 16 bit src_row0_8x16b = _mm_unpacklo_epi8(src_row0_8x16b, tmp2_const_8x16b); src_row1_8x16b = _mm_unpacklo_epi8(src_row1_8x16b, tmp2_const_8x16b); src_row2_8x16b = _mm_unpacklo_epi8(src_row2_8x16b, tmp2_const_8x16b); src_row3_8x16b = _mm_unpacklo_epi8(src_row3_8x16b, tmp2_const_8x16b); //shuffle values loaded tmp0_const_8x16b = _mm_load_si128((__m128i *)shuffle2); tmp1_const_8x16b = _mm_load_si128((__m128i *)shuffle3); //arranging each row delta in different registers tmp_delta3_8x16b = _mm_shuffle_epi8(tmp_delta1_8x16b, tmp1_const_8x16b); tmp_delta2_8x16b = _mm_shuffle_epi8(tmp_delta1_8x16b, tmp0_const_8x16b); tmp_delta1_8x16b = _mm_shuffle_epi8(tmp_delta0_8x16b, tmp1_const_8x16b); tmp_delta0_8x16b = _mm_shuffle_epi8(tmp_delta0_8x16b, tmp0_const_8x16b); //adding the respective delta src_row3_8x16b = _mm_add_epi16(tmp_delta3_8x16b, src_row3_8x16b); src_row2_8x16b = _mm_add_epi16(tmp_delta2_8x16b, src_row2_8x16b); src_row1_8x16b = _mm_add_epi16(tmp_delta1_8x16b, src_row1_8x16b); src_row0_8x16b = _mm_add_epi16(tmp_delta0_8x16b, src_row0_8x16b); //saturating to 8 bit src_row2_8x16b = _mm_packus_epi16(src_row2_8x16b, src_row3_8x16b); src_row0_8x16b = _mm_packus_epi16(src_row0_8x16b, src_row1_8x16b); //separating different rows src_row1_8x16b = _mm_srli_si128(src_row0_8x16b, 8); src_row3_8x16b = _mm_srli_si128(src_row2_8x16b, 8); } _mm_storel_epi64((__m128i *)(pu1_src - 4), src_row0_8x16b); _mm_storel_epi64((__m128i *)((pu1_src - 4) + src_strd), src_row1_8x16b); _mm_storel_epi64((__m128i *)((pu1_src - 4) + 2 * src_strd), src_row2_8x16b); _mm_storel_epi64((__m128i *)((pu1_src - 4) + 3 * src_strd), src_row3_8x16b); } } void ihevc_deblk_luma_horz_ssse3(UWORD8 *pu1_src, WORD32 src_strd, WORD32 bs, WORD32 quant_param_p, WORD32 quant_param_q, WORD32 beta_offset_div2, WORD32 tc_offset_div2, WORD32 filter_flag_p, WORD32 filter_flag_q) { WORD32 qp_luma, beta_indx, tc_indx; WORD32 beta, tc; WORD32 d0, d3, dp, dq, d; WORD32 de_0, de_1, de_2, de_3; WORD32 d_sam0, d_sam3; WORD32 de, dep, deq; __m128i src_q0_8x16b, src_q1_8x16b, src_p0_8x16b, src_p1_8x16b, src_q2_8x16b; __m128i tmp_pq_str1_8x16b, src_p2_8x16b, tmp_pq_str0_8x16b; { __m128i src_tmp_p_0_8x16b, src_tmp_p_1_8x16b, src_tmp_q_0_8x16b, src_tmp_q_1_8x16b; __m128i coef_8x16b, mask_d_result_4x32b, mask_de_result_8x16b; __m128i mask_16x8b, temp_coef0_8x16b, temp_coef1_8x16b; ASSERT((bs > 0)); ASSERT(filter_flag_p || filter_flag_q); qp_luma = (quant_param_p + quant_param_q + 1) >> 1; beta_indx = CLIP3(qp_luma + (beta_offset_div2 << 1), 0, 51); /* BS based on implementation can take value 3 if it is intra/inter egde */ /* based on BS, tc index is calcuated by adding 2 * ( bs - 1) to QP and tc_offset */ /* for BS = 1 adding factor is (0*2), BS = 2 or 3 adding factor is (1*2) */ /* the above desired functionallity is achieved by doing (2*(bs>>1)) */ tc_indx = CLIP3(qp_luma + 2 * (bs >> 1) + (tc_offset_div2 << 1), 0, 53); beta = gai4_ihevc_beta_table[beta_indx]; tc = gai4_ihevc_tc_table[tc_indx]; if(0 == tc) { return; } src_q0_8x16b = _mm_loadl_epi64((__m128i *)(pu1_src)); src_q1_8x16b = _mm_loadl_epi64((__m128i *)(pu1_src + src_strd)); src_p0_8x16b = _mm_loadl_epi64((__m128i *)(pu1_src - src_strd)); src_p1_8x16b = _mm_loadl_epi64((__m128i *)(pu1_src - 2 * src_strd)); src_q2_8x16b = _mm_loadl_epi64((__m128i *)(pu1_src + 2 * src_strd)); tmp_pq_str1_8x16b = _mm_loadl_epi64((__m128i *)(pu1_src + 3 * src_strd)); src_p2_8x16b = _mm_loadl_epi64((__m128i *)(pu1_src - 3 * src_strd)); tmp_pq_str0_8x16b = _mm_loadl_epi64((__m128i *)(pu1_src - 4 * src_strd)); src_tmp_p_0_8x16b = _mm_unpacklo_epi8(src_p1_8x16b, src_p0_8x16b); src_tmp_p_1_8x16b = _mm_unpacklo_epi8(tmp_pq_str0_8x16b, src_p2_8x16b); src_tmp_q_0_8x16b = _mm_unpacklo_epi8(src_q0_8x16b, src_q1_8x16b); src_tmp_q_1_8x16b = _mm_unpacklo_epi8(src_q2_8x16b, tmp_pq_str1_8x16b); src_tmp_p_0_8x16b = _mm_unpacklo_epi16(src_tmp_p_1_8x16b, src_tmp_p_0_8x16b); src_tmp_q_0_8x16b = _mm_unpacklo_epi16(src_tmp_q_0_8x16b, src_tmp_q_1_8x16b); src_tmp_p_0_8x16b = _mm_shuffle_epi32(src_tmp_p_0_8x16b, 0x6c); src_tmp_q_0_8x16b = _mm_shuffle_epi32(src_tmp_q_0_8x16b, 0x6c); coef_8x16b = _mm_load_si128((__m128i *)(coef_d)); mask_16x8b = _mm_load_si128((__m128i *)(shuffle_d)); src_tmp_p_0_8x16b = _mm_unpacklo_epi32(src_tmp_p_0_8x16b, src_tmp_q_0_8x16b); //WORD32 shuffle_d[4]={0x80800403,0x80800c0b,0x03000704,0x0b080f0c}; mask_de_result_8x16b = _mm_shuffle_epi8(src_tmp_p_0_8x16b, mask_16x8b); mask_d_result_4x32b = _mm_maddubs_epi16(src_tmp_p_0_8x16b, coef_8x16b); //to get all 1's of 8 bit in (1) temp_coef0_8x16b = _mm_cmpeq_epi16(src_tmp_p_0_8x16b, src_tmp_p_0_8x16b); temp_coef1_8x16b = _mm_srli_epi16(temp_coef0_8x16b, 15); //accumulating values foe dp3 dq3 , dp0 dq0 values mask_d_result_4x32b = _mm_madd_epi16(mask_d_result_4x32b, temp_coef1_8x16b); temp_coef1_8x16b = _mm_packus_epi16(temp_coef1_8x16b, temp_coef1_8x16b); // to get all 1,-1 sets of 16 bits in (0) temp_coef0_8x16b = _mm_unpacklo_epi8(temp_coef0_8x16b, temp_coef1_8x16b); //q33-q30,p33-p30,q03-q00,p03-p00,0,q30-p30,0,q00-p00 mask_de_result_8x16b = _mm_maddubs_epi16(mask_de_result_8x16b, temp_coef0_8x16b); //to get 16 bit 1's temp_coef0_8x16b = _mm_srli_epi16(temp_coef1_8x16b, 8); // dq3 dp3 dq0 dp0 mask_d_result_4x32b = _mm_abs_epi32(mask_d_result_4x32b); mask_16x8b = _mm_shuffle_epi32(mask_d_result_4x32b, 0xec); mask_d_result_4x32b = _mm_shuffle_epi32(mask_d_result_4x32b, 0x49); // dq dp d3 d0 mask_d_result_4x32b = _mm_add_epi32(mask_d_result_4x32b, mask_16x8b); //|q33-q30|,|p33-p30|,|q03-q00|,|p03-p00|,0,|q30-p30|,0,|q00-p00| mask_de_result_8x16b = _mm_abs_epi16(mask_de_result_8x16b); //|q33-q30|+|p33-p30|,|q03-q00|+|p03-p00|,0+|q30-p30|,0+|q00-p00| mask_de_result_8x16b = _mm_madd_epi16(mask_de_result_8x16b, temp_coef0_8x16b); ///store back in a single variable temp_coef0_8x16b = _mm_srli_si128(mask_d_result_4x32b, 4); temp_coef1_8x16b = _mm_srli_si128(mask_d_result_4x32b, 8); mask_16x8b = _mm_srli_si128(mask_d_result_4x32b, 12); d0 = _mm_cvtsi128_si32(mask_d_result_4x32b); d3 = _mm_cvtsi128_si32(temp_coef0_8x16b); dp = _mm_cvtsi128_si32(temp_coef1_8x16b); dq = _mm_cvtsi128_si32(mask_16x8b); //getting d d = d0 + d3; ///store back in a single variable temp_coef0_8x16b = _mm_srli_si128(mask_de_result_8x16b, 4); temp_coef1_8x16b = _mm_srli_si128(mask_de_result_8x16b, 8); mask_16x8b = _mm_srli_si128(mask_de_result_8x16b, 12); de_0 = _mm_cvtsi128_si32(mask_de_result_8x16b); de_1 = _mm_cvtsi128_si32(temp_coef0_8x16b); de_2 = _mm_cvtsi128_si32(temp_coef1_8x16b); de_3 = _mm_cvtsi128_si32(mask_16x8b); de = 0; dep = 0; deq = 0; if(d < beta) { d_sam0 = 0; if((2 * d0 < (beta >> 2)) && (de_2 < (beta >> 3)) && (de_0 < ((5 * tc + 1) >> 1))) { d_sam0 = 1; } d_sam3 = 0; if((2 * d3 < (beta >> 2)) && (de_3 < (beta >> 3)) && de_1 < ((5 * tc + 1) >> 1)) { d_sam3 = 1; } de = (d_sam0 & d_sam3) + 1; dep = (dp < (beta + (beta >> 1)) >> 3) ? 1 : 0; deq = (dq < (beta + (beta >> 1)) >> 3) ? 1 : 0; if(tc <= 1) { dep = 0; deq = 0; } } } if(de != 0) { if(2 == de) { __m128i temp_pq0_str0_16x8b; __m128i temp_pq1_str0_16x8b, temp_pq1_str1_16x8b; __m128i temp_pq2_str0_16x8b; __m128i temp_str0_16x8b, temp_str1_16x8b; __m128i const2_8x16b, const2tc_8x16b; LWORD64 mask, tc2; tc = tc << 1; mask = (((LWORD64)filter_flag_q) << 63) | (((LWORD64)filter_flag_p) << 31); tc2 = ((LWORD64)tc); const2_8x16b = _mm_cmpeq_epi16(src_p1_8x16b, src_p1_8x16b); //q'0-q'1-2 ,p'0-p'1-2 temp_pq0_str0_16x8b = _mm_unpacklo_epi8(src_p1_8x16b, src_p0_8x16b); temp_str0_16x8b = _mm_unpacklo_epi8(src_q0_8x16b, src_q1_8x16b); const2_8x16b = _mm_srli_epi16(const2_8x16b, 15); //arranged q31 q30 q21 q20 q1 q10 q01 q00 p30 p31 p20 p21 p10 p11 p00 p01 temp_pq0_str0_16x8b = _mm_unpacklo_epi64(temp_pq0_str0_16x8b, temp_str0_16x8b); const2_8x16b = _mm_packus_epi16(const2_8x16b, const2_8x16b); temp_pq0_str0_16x8b = _mm_maddubs_epi16(temp_pq0_str0_16x8b, const2_8x16b); //q'1-2, p'1-2 temp_pq1_str0_16x8b = _mm_unpacklo_epi8(src_p0_8x16b, src_q0_8x16b); temp_pq1_str1_16x8b = _mm_unpacklo_epi8(src_q1_8x16b, src_q2_8x16b); temp_str1_16x8b = _mm_unpacklo_epi8(src_p1_8x16b, src_p2_8x16b); // q30 p30 q20 p20 q10 p10 q01 q00 p30 q20 p20 q10 p10 q01 q00 p00 temp_pq1_str0_16x8b = _mm_unpacklo_epi64(temp_pq1_str0_16x8b, temp_pq1_str0_16x8b); // q32 q31 q22 q21 q12 q11 q02 q01 p32 p31 p22 p21 p12 p11 p02 p01 temp_pq1_str1_16x8b = _mm_unpacklo_epi64(temp_str1_16x8b, temp_pq1_str1_16x8b); temp_pq1_str0_16x8b = _mm_maddubs_epi16(temp_pq1_str0_16x8b, const2_8x16b); temp_pq1_str1_16x8b = _mm_maddubs_epi16(temp_pq1_str1_16x8b, const2_8x16b); //clipping mask design temp_str1_16x8b = _mm_setzero_si128(); temp_str0_16x8b = _mm_loadl_epi64((__m128i *)(&mask)); const2tc_8x16b = _mm_loadl_epi64((__m128i *)(&tc2)); temp_str0_16x8b = _mm_shuffle_epi32(temp_str0_16x8b, 0x44); const2tc_8x16b = _mm_shuffle_epi8(const2tc_8x16b, temp_str1_16x8b); //clipping mask design temp_str0_16x8b = _mm_srai_epi32(temp_str0_16x8b, 31); const2tc_8x16b = _mm_and_si128(const2tc_8x16b, temp_str0_16x8b); //calculating Clipping MAX for all pixel values. src_p0_8x16b = _mm_unpacklo_epi32(src_p0_8x16b, src_q0_8x16b); src_q0_8x16b = _mm_unpacklo_epi32(src_p1_8x16b, src_q1_8x16b); //for clipping calc src_p1_8x16b = _mm_unpacklo_epi64(src_p0_8x16b, src_q0_8x16b); //saving the unmodified data of q1 p1 q0 p0 src_q1_8x16b = _mm_unpackhi_epi64(src_p0_8x16b, src_q0_8x16b); //CLIpping MAX and MIN for q1 p1 q0 p0 src_p0_8x16b = _mm_adds_epu8(src_p1_8x16b, const2tc_8x16b); src_p1_8x16b = _mm_subs_epu8(src_p1_8x16b, const2tc_8x16b); //q'2-q'0-2,p'2-p'0-2 tmp_pq_str0_8x16b = _mm_unpacklo_epi8(src_p2_8x16b, tmp_pq_str0_8x16b); temp_pq2_str0_16x8b = _mm_unpacklo_epi8(src_q2_8x16b, tmp_pq_str1_8x16b); const2_8x16b = _mm_slli_epi16(const2_8x16b, 1); //arranged q33 q32 q23 q22 q13 q12 q03 q02 p32 p33 p22 p23 p12 p13 p02 p03 temp_pq2_str0_16x8b = _mm_unpacklo_epi64(tmp_pq_str0_8x16b, temp_pq2_str0_16x8b); src_p2_8x16b = _mm_unpacklo_epi32(src_p2_8x16b, src_q2_8x16b); temp_pq2_str0_16x8b = _mm_maddubs_epi16(temp_pq2_str0_16x8b, const2_8x16b); //calculating Clipping MAX and MIN for p2 and q2 . tmp_pq_str0_8x16b = _mm_adds_epu8(src_p2_8x16b, const2tc_8x16b); tmp_pq_str1_8x16b = _mm_subs_epu8(src_p2_8x16b, const2tc_8x16b); //q'0-q'1-2 ,p'0-p'1-2 temp_str0_16x8b = _mm_shuffle_epi32(temp_pq0_str0_16x8b, 0x4e); temp_pq0_str0_16x8b = _mm_add_epi16(temp_pq0_str0_16x8b, temp_str0_16x8b); //q'1-2 p'1-2 temp_pq1_str0_16x8b = _mm_add_epi16(temp_pq1_str0_16x8b, temp_pq1_str1_16x8b); //to get 2 in 16 bit const2_8x16b = _mm_srli_epi16(const2_8x16b, 8); //q'1, p'1 (adding 2) temp_pq1_str0_16x8b = _mm_add_epi16(temp_pq1_str0_16x8b, const2_8x16b); //q'0-q'1,p'0-p'1 temp_pq0_str0_16x8b = _mm_add_epi16(temp_pq0_str0_16x8b, const2_8x16b); //q'2-q'1,p'2-p'1 temp_pq2_str0_16x8b = _mm_add_epi16(temp_pq2_str0_16x8b, const2_8x16b); //q'0 = (q'0-q'1)+q'1 ,p'0 = (p'0-p'1)+p'1; temp_pq0_str0_16x8b = _mm_add_epi16(temp_pq1_str0_16x8b, temp_pq0_str0_16x8b); //q'2 = (q'2-q'1)+q'1 ,p'2 = (p'2-p'1)+p'1; temp_pq2_str0_16x8b = _mm_add_epi16(temp_pq1_str0_16x8b, temp_pq2_str0_16x8b); //normalisation of all modified pixels temp_pq0_str0_16x8b = _mm_srai_epi16(temp_pq0_str0_16x8b, 3); temp_pq1_str0_16x8b = _mm_srai_epi16(temp_pq1_str0_16x8b, 2); temp_pq2_str0_16x8b = _mm_srai_epi16(temp_pq2_str0_16x8b, 3); //q'1 p'1 q'0 p'0 temp_pq0_str0_16x8b = _mm_packus_epi16(temp_pq0_str0_16x8b, temp_pq1_str0_16x8b); temp_pq2_str0_16x8b = _mm_packus_epi16(temp_pq2_str0_16x8b, temp_pq2_str0_16x8b); //pack with the unmodified data of q2 and p2 src_p2_8x16b = _mm_unpackhi_epi64(temp_pq2_str0_16x8b, src_p2_8x16b); //Clipping MAX and MIN for q'1 p'1 q'0 p'0 and q'2 p'2 temp_pq0_str0_16x8b = _mm_min_epu8(temp_pq0_str0_16x8b, src_p0_8x16b); src_p2_8x16b = _mm_min_epu8(src_p2_8x16b, tmp_pq_str0_8x16b); temp_pq0_str0_16x8b = _mm_max_epu8(temp_pq0_str0_16x8b, src_p1_8x16b); src_p2_8x16b = _mm_max_epu8(src_p2_8x16b, tmp_pq_str1_8x16b); //Reshuffling q'1 p'1 q'0 p'0 along with unmodified data src_p0_8x16b = _mm_unpacklo_epi32(temp_pq0_str0_16x8b, src_q1_8x16b); src_p1_8x16b = _mm_unpackhi_epi32(temp_pq0_str0_16x8b, src_q1_8x16b); src_p2_8x16b = _mm_shuffle_epi32(src_p2_8x16b, 0xd8); src_q0_8x16b = _mm_srli_si128(src_p0_8x16b, 8); src_q1_8x16b = _mm_srli_si128(src_p1_8x16b, 8); src_q2_8x16b = _mm_srli_si128(src_p2_8x16b, 8); _mm_storel_epi64((__m128i *)(pu1_src - 3 * src_strd), src_p2_8x16b); _mm_storel_epi64((__m128i *)(pu1_src - 2 * src_strd), src_p1_8x16b); _mm_storel_epi64((__m128i *)(pu1_src - src_strd), src_p0_8x16b); _mm_storel_epi64((__m128i *)(pu1_src), src_q0_8x16b); _mm_storel_epi64((__m128i *)(pu1_src + src_strd), src_q1_8x16b); _mm_storel_epi64((__m128i *)(pu1_src + 2 * src_strd), src_q2_8x16b); } else { __m128i tmp_delta0_8x16b, tmp_delta1_8x16b; __m128i tmp0_const_8x16b, tmp1_const_8x16b, tmp2_const_8x16b; __m128i coefdelta_0_8x16b; __m128i const2_8x16b, consttc_8x16b; LWORD64 maskp0, maskp1, maskq0, maskq1; maskp0 = (LWORD64)filter_flag_p; maskq0 = (LWORD64)filter_flag_q; maskp1 = (LWORD64)dep; maskq1 = (LWORD64)deq; consttc_8x16b = _mm_set1_epi32(tc); tmp_delta0_8x16b = _mm_unpacklo_epi8(src_p1_8x16b, src_p0_8x16b); tmp_delta1_8x16b = _mm_unpacklo_epi8(src_q0_8x16b, src_q1_8x16b); //arranged q31 q30 p30 p31 q21 q20 p20 p21 q1 q10 p10 p11 q01 q00 p00 p01 tmp_delta1_8x16b = _mm_unpacklo_epi16(tmp_delta0_8x16b, tmp_delta1_8x16b); coefdelta_0_8x16b = _mm_load_si128((__m128i *)coef_de1); // (-3q1+9q0),(-9p0+3p1) tmp_delta0_8x16b = _mm_maddubs_epi16(tmp_delta1_8x16b, coefdelta_0_8x16b); //getting -tc store tmp2_const_8x16b = _mm_cmpeq_epi32(consttc_8x16b, consttc_8x16b); //getting tc in 16 bit consttc_8x16b = _mm_packs_epi32(consttc_8x16b, consttc_8x16b); //calc 10 *tc = 2*tc +8*tc ; 2*tc tmp_pq_str0_8x16b = _mm_slli_epi16(consttc_8x16b, 1); //calc 10 *tc = 2*tc +8*tc ; 8*tc tmp_pq_str1_8x16b = _mm_slli_epi16(consttc_8x16b, 3); //const 1 const2_8x16b = _mm_srli_epi16(tmp2_const_8x16b, 15); //calc 10 *tc tmp_pq_str0_8x16b = _mm_add_epi16(tmp_pq_str0_8x16b, tmp_pq_str1_8x16b); //delta0 without normalisation and clipping tmp_delta0_8x16b = _mm_madd_epi16(tmp_delta0_8x16b, const2_8x16b); const2_8x16b = _mm_srli_epi32(tmp2_const_8x16b, 31); //loaded coef for delta1 calculation coefdelta_0_8x16b = _mm_load_si128((__m128i *)coef_dep1); //(-2q1+q0),(p0-2p1) tmp_delta1_8x16b = _mm_maddubs_epi16(tmp_delta1_8x16b, coefdelta_0_8x16b); //const 8 const2_8x16b = _mm_slli_epi32(const2_8x16b, 3); //normalisation of the filter tmp_delta0_8x16b = _mm_add_epi32(tmp_delta0_8x16b, const2_8x16b); tmp_delta0_8x16b = _mm_srai_epi32(tmp_delta0_8x16b, 4); //getting deltaq0 tmp_pq_str1_8x16b = _mm_sign_epi32(tmp_delta0_8x16b, tmp2_const_8x16b); //getting -tc tmp1_const_8x16b = _mm_sign_epi16(consttc_8x16b, tmp2_const_8x16b); //packing d03q d02q d01q d0q d03p d02p d01p d00p tmp_delta0_8x16b = _mm_packs_epi32(tmp_delta0_8x16b, tmp_pq_str1_8x16b); //absolute delta tmp_pq_str1_8x16b = _mm_abs_epi16(tmp_delta0_8x16b); //Clipping of delta0 tmp_delta0_8x16b = _mm_min_epi16(tmp_delta0_8x16b, consttc_8x16b); //tc>>1 16 bit consttc_8x16b = _mm_srai_epi16(consttc_8x16b, 1); //Clipping of delta0 tmp_delta0_8x16b = _mm_max_epi16(tmp_delta0_8x16b, tmp1_const_8x16b); //(-tc)>>1 16 bit tmp1_const_8x16b = _mm_sign_epi16(consttc_8x16b, tmp2_const_8x16b); //mask for |delta| < 10*tc tmp_pq_str0_8x16b = _mm_cmpgt_epi16(tmp_pq_str0_8x16b, tmp_pq_str1_8x16b); //delta 1 calc starts //getting q32 q22 q12 q02 p32 p12 p22 p02 tmp0_const_8x16b = _mm_setzero_si128(); src_q2_8x16b = _mm_unpacklo_epi8(src_q2_8x16b, tmp0_const_8x16b); src_p2_8x16b = _mm_unpacklo_epi8(src_p2_8x16b, tmp0_const_8x16b); src_p2_8x16b = _mm_unpacklo_epi64(src_p2_8x16b, src_q2_8x16b); //constant 1 const2_8x16b = _mm_srli_epi16(tmp2_const_8x16b, 15); //2*delta0 tmp2_const_8x16b = _mm_add_epi16(tmp_delta0_8x16b, tmp_delta0_8x16b); //getting all respective q's and p's together coefdelta_0_8x16b = _mm_load_si128((__m128i *)(shuffle1)); tmp_delta1_8x16b = _mm_shuffle_epi8(tmp_delta1_8x16b, coefdelta_0_8x16b); //final adds for deltap1 and deltaq1 tmp_delta1_8x16b = _mm_add_epi16(tmp_delta1_8x16b, const2_8x16b); src_p2_8x16b = _mm_add_epi16(src_p2_8x16b, tmp2_const_8x16b); tmp_delta1_8x16b = _mm_add_epi16(tmp_delta1_8x16b, src_p2_8x16b); tmp_delta1_8x16b = _mm_srai_epi16(tmp_delta1_8x16b, 2); //mask0= (((LWORD64)filter_flag_q)<<63)| (((LWORD64)filter_flag_p)<<31); tmp_pq_str1_8x16b = _mm_loadl_epi64((__m128i *)(&(maskq0))); src_p2_8x16b = _mm_loadl_epi64((__m128i *)(&(maskp0))); // src_p2_8x16b = _mm_set_epi32(filter_flag_q,filter_flag_p,filter_flag_q,filter_flag_p); //mask1= (((LWORD64)(filter_flag_q&deq))<<63)|(((LWORD64)(filter_flag_p & dep))<<31); src_q2_8x16b = _mm_loadl_epi64((__m128i *)(&(maskq1))); coefdelta_0_8x16b = _mm_loadl_epi64((__m128i *)(&(maskp1))); src_p2_8x16b = _mm_unpacklo_epi32(src_p2_8x16b, tmp_pq_str1_8x16b); src_q2_8x16b = _mm_unpacklo_epi32(coefdelta_0_8x16b, src_q2_8x16b); //src_q2_8x16b = _mm_set_epi32(deq,dep,deq,dep); src_q2_8x16b = _mm_and_si128(src_q2_8x16b, src_p2_8x16b); //rearranging the mask values src_q2_8x16b = _mm_shuffle_epi32(src_q2_8x16b, 0x50); src_p2_8x16b = _mm_shuffle_epi32(src_p2_8x16b, 0x50); src_q2_8x16b = _mm_slli_epi32(src_q2_8x16b, 31); src_p2_8x16b = _mm_slli_epi32(src_p2_8x16b, 31); src_q2_8x16b = _mm_srai_epi32(src_q2_8x16b, 31); src_p2_8x16b = _mm_srai_epi32(src_p2_8x16b, 31); //combining mask delta1 tmp_pq_str1_8x16b = _mm_and_si128(tmp_pq_str0_8x16b, src_q2_8x16b); // clipping delta1 tmp_delta1_8x16b = _mm_min_epi16(tmp_delta1_8x16b, consttc_8x16b); //combining mask delat0 tmp_pq_str0_8x16b = _mm_and_si128(tmp_pq_str0_8x16b, src_p2_8x16b); // clipping delta1 tmp_delta1_8x16b = _mm_max_epi16(tmp_delta1_8x16b, tmp1_const_8x16b); //masking of the delta values |delta|<10*tc tmp_delta1_8x16b = _mm_and_si128(tmp_delta1_8x16b, tmp_pq_str1_8x16b); tmp_delta0_8x16b = _mm_and_si128(tmp_delta0_8x16b, tmp_pq_str0_8x16b); //separating p and q delta 0 and addinq p0 and q0 tmp_pq_str0_8x16b = _mm_unpacklo_epi64(tmp_delta0_8x16b, tmp0_const_8x16b); tmp_pq_str1_8x16b = _mm_unpackhi_epi64(tmp_delta0_8x16b, tmp0_const_8x16b); src_p0_8x16b = _mm_unpacklo_epi8(src_p0_8x16b, tmp0_const_8x16b); src_q0_8x16b = _mm_unpacklo_epi8(src_q0_8x16b, tmp0_const_8x16b); src_p0_8x16b = _mm_add_epi16(src_p0_8x16b, tmp_pq_str0_8x16b); src_q0_8x16b = _mm_add_epi16(src_q0_8x16b, tmp_pq_str1_8x16b); //separating p and q delta 0 and addinq p0 and q0 tmp_pq_str0_8x16b = _mm_unpacklo_epi64(tmp_delta1_8x16b, tmp0_const_8x16b); tmp_pq_str1_8x16b = _mm_unpackhi_epi64(tmp_delta1_8x16b, tmp0_const_8x16b); src_p1_8x16b = _mm_unpacklo_epi8(src_p1_8x16b, tmp0_const_8x16b); src_q1_8x16b = _mm_unpacklo_epi8(src_q1_8x16b, tmp0_const_8x16b); src_p1_8x16b = _mm_add_epi16(src_p1_8x16b, tmp_pq_str0_8x16b); src_q1_8x16b = _mm_add_epi16(src_q1_8x16b, tmp_pq_str1_8x16b); //packing p1 q1 and p0 q0 to 8 bit src_p1_8x16b = _mm_packus_epi16(src_p1_8x16b, src_q1_8x16b); src_p0_8x16b = _mm_packus_epi16(src_p0_8x16b, src_q0_8x16b); src_q1_8x16b = _mm_srli_si128(src_p1_8x16b, 8); src_q0_8x16b = _mm_srli_si128(src_p0_8x16b, 8); _mm_storel_epi64((__m128i *)(pu1_src - 2 * src_strd), src_p1_8x16b); _mm_storel_epi64((__m128i *)(pu1_src - src_strd), src_p0_8x16b); _mm_storel_epi64((__m128i *)(pu1_src), src_q0_8x16b); _mm_storel_epi64((__m128i *)(pu1_src + src_strd), src_q1_8x16b); } } } void ihevc_deblk_chroma_vert_ssse3(UWORD8 *pu1_src, WORD32 src_strd, WORD32 quant_param_p, WORD32 quant_param_q, WORD32 qp_offset_u, WORD32 qp_offset_v, WORD32 tc_offset_div2, WORD32 filter_flag_p, WORD32 filter_flag_q) { WORD32 qp_indx_u, qp_chroma_u; WORD32 qp_indx_v, qp_chroma_v; WORD32 tc_indx_u, tc_u; WORD32 tc_indx_v, tc_v; __m128i src_row_0_16x8b, tmp_pxl_0_16x8b, src_row_2_16x8b, tmp_pxl_1_16x8b; ASSERT(filter_flag_p || filter_flag_q); /* chroma processing is done only if BS is 2 */ /* this function is assumed to be called only if BS is 2 */ qp_indx_u = qp_offset_u + ((quant_param_p + quant_param_q + 1) >> 1); qp_chroma_u = qp_indx_u < 0 ? qp_indx_u : (qp_indx_u > 57 ? qp_indx_u - 6 : gai4_ihevc_qp_table[qp_indx_u]); qp_indx_v = qp_offset_v + ((quant_param_p + quant_param_q + 1) >> 1); qp_chroma_v = qp_indx_v < 0 ? qp_indx_v : (qp_indx_v > 57 ? qp_indx_v - 6 : gai4_ihevc_qp_table[qp_indx_v]); tc_indx_u = CLIP3(qp_chroma_u + 2 + (tc_offset_div2 << 1), 0, 53); tc_u = gai4_ihevc_tc_table[tc_indx_u]; tc_indx_v = CLIP3(qp_chroma_v + 2 + (tc_offset_div2 << 1), 0, 53); tc_v = gai4_ihevc_tc_table[tc_indx_v]; if(0 == tc_u && 0 == tc_v) { return; } src_row_0_16x8b = _mm_loadl_epi64((__m128i *)(pu1_src - 4)); tmp_pxl_0_16x8b = _mm_loadl_epi64((__m128i *)(pu1_src + src_strd - 4)); src_row_2_16x8b = _mm_loadl_epi64((__m128i *)(pu1_src + 2 * src_strd - 4)); tmp_pxl_1_16x8b = _mm_loadl_epi64((__m128i *)(pu1_src + 3 * src_strd - 4)); { LWORD64 mask_tc, mask_flag, mask; __m128i delta_vu0_16x8b, delta_vu1_16x8b; __m128i mask_tc_16x8, mask_16x8b, mask_flag_p_16x8b, mask_flag_q_16x8b; __m128i min_0_16x8b; __m128i const_16x8b; mask_flag = (((LWORD64)filter_flag_p) << 31) | (((LWORD64)filter_flag_q) << 63); mask_tc = (((LWORD64)tc_v) << 16) | ((LWORD64)tc_u); mask = 0xffff00000000ffffLL; src_row_0_16x8b = _mm_unpacklo_epi64(src_row_0_16x8b, tmp_pxl_0_16x8b); src_row_2_16x8b = _mm_unpacklo_epi64(src_row_2_16x8b, tmp_pxl_1_16x8b); mask_16x8b = _mm_load_si128((__m128i *)(shuffle_uv)); // qv11 qu11 qv10 qu10 qv01 qu01 qv00 qu00 pv10 pu10 pv11 pu11 pv00 pu00 pv01 pu01 // qv31 qu31 qv30 qu30 qv21 qu21 qv20 qu20 pv30 pu30 pv31 pu31 pv20 pu20 pv21 pu21 delta_vu0_16x8b = _mm_shuffle_epi8(src_row_0_16x8b, mask_16x8b); delta_vu1_16x8b = _mm_shuffle_epi8(src_row_2_16x8b, mask_16x8b); tmp_pxl_0_16x8b = _mm_unpacklo_epi64(delta_vu0_16x8b, delta_vu1_16x8b); tmp_pxl_1_16x8b = _mm_unpackhi_epi64(delta_vu0_16x8b, delta_vu1_16x8b); // pv30 pv31 pu30 pu31 pv20 pv21 pu20 pu21 pv10 pv11 pu10 pu11 pv00 pv01 pu00 pu01 // qv31 qv30 qu31 qu30 qv21 qv20 qu21 qu20 qv11 qv10 qu11 qu10 qv01 qv00 qu01 qu00 delta_vu0_16x8b = _mm_load_si128((__m128i *)delta0); delta_vu1_16x8b = _mm_load_si128((__m128i *)delta1); delta_vu0_16x8b = _mm_maddubs_epi16(tmp_pxl_0_16x8b, delta_vu0_16x8b); delta_vu1_16x8b = _mm_maddubs_epi16(tmp_pxl_1_16x8b, delta_vu1_16x8b); //generating offset 4 const_16x8b = _mm_cmpeq_epi16(tmp_pxl_0_16x8b, tmp_pxl_0_16x8b); // filter flag mask and tc mask mask_tc_16x8 = _mm_loadl_epi64((__m128i *)(&mask_tc)); mask_flag_q_16x8b = _mm_loadl_epi64((__m128i *)(&mask_flag)); mask_tc_16x8 = _mm_shuffle_epi32(mask_tc_16x8, 0x00); mask_flag_q_16x8b = _mm_srai_epi32(mask_flag_q_16x8b, 31); //-tc min_0_16x8b = _mm_sign_epi16(mask_tc_16x8, const_16x8b); //converting const 1 const_16x8b = _mm_srli_epi16(const_16x8b, 15); //filterp and filterq flag mask_flag_p_16x8b = _mm_shuffle_epi32(mask_flag_q_16x8b, 0x00); mask_flag_q_16x8b = _mm_shuffle_epi32(mask_flag_q_16x8b, 0x55); //modified delta with a filter (1 -4 4 -1) available in 16 bit delta_vu0_16x8b = _mm_add_epi16(delta_vu0_16x8b, delta_vu1_16x8b); //converting const 4 const_16x8b = _mm_slli_epi16(const_16x8b, 2); mask_16x8b = _mm_loadl_epi64((__m128i *)(&mask)); //offset addition delta_vu0_16x8b = _mm_add_epi16(delta_vu0_16x8b, const_16x8b); //eliminating q1 tmp_pxl_1_16x8b = _mm_slli_epi16(tmp_pxl_1_16x8b, 8); const_16x8b = _mm_setzero_si128(); //filter after normalisation delta_vu0_16x8b = _mm_srai_epi16(delta_vu0_16x8b, 3); mask_16x8b = _mm_shuffle_epi32(mask_16x8b, 0x44); //clipping MAX delta_vu0_16x8b = _mm_min_epi16(delta_vu0_16x8b, mask_tc_16x8); //getting p0 and eliminating p1 tmp_pxl_0_16x8b = _mm_srli_epi16(tmp_pxl_0_16x8b, 8); //clipping MIN delta_vu0_16x8b = _mm_max_epi16(delta_vu0_16x8b, min_0_16x8b); //getting q0 tmp_pxl_1_16x8b = _mm_srli_epi16(tmp_pxl_1_16x8b, 8); //masking filter flag delta_vu1_16x8b = _mm_and_si128(delta_vu0_16x8b, mask_flag_q_16x8b); delta_vu0_16x8b = _mm_and_si128(delta_vu0_16x8b, mask_flag_p_16x8b); // q-delta ,p+delta tmp_pxl_1_16x8b = _mm_sub_epi16(tmp_pxl_1_16x8b, delta_vu1_16x8b); tmp_pxl_0_16x8b = _mm_add_epi16(tmp_pxl_0_16x8b, delta_vu0_16x8b); //merging q0 and p0 of respective rows delta_vu1_16x8b = _mm_unpackhi_epi32(tmp_pxl_0_16x8b, tmp_pxl_1_16x8b); delta_vu0_16x8b = _mm_unpacklo_epi32(tmp_pxl_0_16x8b, tmp_pxl_1_16x8b); // row 0 and row 1 packed , row2 and row3 packed delta_vu0_16x8b = _mm_packus_epi16(delta_vu0_16x8b, const_16x8b); delta_vu1_16x8b = _mm_packus_epi16(delta_vu1_16x8b, const_16x8b); //removing older pixel values src_row_0_16x8b = _mm_and_si128(src_row_0_16x8b, mask_16x8b); src_row_2_16x8b = _mm_and_si128(src_row_2_16x8b, mask_16x8b); //arranging modified pixels delta_vu0_16x8b = _mm_shuffle_epi32(delta_vu0_16x8b, 0xd8); delta_vu1_16x8b = _mm_shuffle_epi32(delta_vu1_16x8b, 0xd8); delta_vu0_16x8b = _mm_slli_epi64(delta_vu0_16x8b, 16); delta_vu1_16x8b = _mm_slli_epi64(delta_vu1_16x8b, 16); //plugging the modified values src_row_0_16x8b = _mm_or_si128(src_row_0_16x8b, delta_vu0_16x8b); src_row_2_16x8b = _mm_or_si128(src_row_2_16x8b, delta_vu1_16x8b); //geting values for row1 and row 3 tmp_pxl_0_16x8b = _mm_srli_si128(src_row_0_16x8b, 8); tmp_pxl_1_16x8b = _mm_srli_si128(src_row_2_16x8b, 8); _mm_storel_epi64((__m128i *)(pu1_src - 4), src_row_0_16x8b); _mm_storel_epi64((__m128i *)((pu1_src - 4) + src_strd), tmp_pxl_0_16x8b); _mm_storel_epi64((__m128i *)((pu1_src - 4) + 2 * src_strd), src_row_2_16x8b); _mm_storel_epi64((__m128i *)((pu1_src - 4) + 3 * src_strd), tmp_pxl_1_16x8b); } } void ihevc_deblk_chroma_horz_ssse3(UWORD8 *pu1_src, WORD32 src_strd, WORD32 quant_param_p, WORD32 quant_param_q, WORD32 qp_offset_u, WORD32 qp_offset_v, WORD32 tc_offset_div2, WORD32 filter_flag_p, WORD32 filter_flag_q) { WORD32 qp_indx_u, qp_chroma_u; WORD32 qp_indx_v, qp_chroma_v; WORD32 tc_indx_u, tc_u; WORD32 tc_indx_v, tc_v; __m128i tmp_p0_16x8b, src_p0_16x8b, src_q0_16x8b, tmp_q0_16x8b; ASSERT(filter_flag_p || filter_flag_q); /* chroma processing is done only if BS is 2 */ /* this function is assumed to be called only if BS is 2 */ qp_indx_u = qp_offset_u + ((quant_param_p + quant_param_q + 1) >> 1); qp_chroma_u = qp_indx_u < 0 ? qp_indx_u : (qp_indx_u > 57 ? qp_indx_u - 6 : gai4_ihevc_qp_table[qp_indx_u]); qp_indx_v = qp_offset_v + ((quant_param_p + quant_param_q + 1) >> 1); qp_chroma_v = qp_indx_v < 0 ? qp_indx_v : (qp_indx_v > 57 ? qp_indx_v - 6 : gai4_ihevc_qp_table[qp_indx_v]); tc_indx_u = CLIP3(qp_chroma_u + 2 + (tc_offset_div2 << 1), 0, 53); tc_u = gai4_ihevc_tc_table[tc_indx_u]; tc_indx_v = CLIP3(qp_chroma_v + 2 + (tc_offset_div2 << 1), 0, 53); tc_v = gai4_ihevc_tc_table[tc_indx_v]; if(0 == tc_u && 0 == tc_v) { return; } tmp_p0_16x8b = _mm_loadl_epi64((__m128i *)(pu1_src - 2 * src_strd)); src_p0_16x8b = _mm_loadl_epi64((__m128i *)(pu1_src - src_strd)); src_q0_16x8b = _mm_loadl_epi64((__m128i *)(pu1_src)); tmp_q0_16x8b = _mm_loadl_epi64((__m128i *)(pu1_src + src_strd)); { LWORD64 mask_tc, mask_flag; __m128i delta_vu0_16x8b, delta_vu1_16x8b; __m128i mask_tc_16x8, mask_16x8b, mask_flag_p_16x8b, mask_flag_q_16x8b; __m128i min_0_16x8b; __m128i const_16x8b; mask_flag = (((LWORD64)filter_flag_p) << 31) | (((LWORD64)filter_flag_q) << 63); mask_tc = (((LWORD64)tc_v) << 16) | ((LWORD64)tc_u); tmp_p0_16x8b = _mm_unpacklo_epi8(tmp_p0_16x8b, src_p0_16x8b); tmp_q0_16x8b = _mm_unpacklo_epi8(src_q0_16x8b, tmp_q0_16x8b); // pv30 pv31 pu30 pu31 pv20 pv21 pu20 pu21 pv10 pv11 pu10 pu11 pv00 pv01 pu00 pu01 // qv31 qv30 qu31 qu30 qv21 qv20 qu21 qu20 qv11 qv10 qu11 qu10 qv01 qv00 qu01 qu00 delta_vu0_16x8b = _mm_load_si128((__m128i *)delta0); delta_vu1_16x8b = _mm_load_si128((__m128i *)delta1); delta_vu0_16x8b = _mm_maddubs_epi16(tmp_p0_16x8b, delta_vu0_16x8b); delta_vu1_16x8b = _mm_maddubs_epi16(tmp_q0_16x8b, delta_vu1_16x8b); // filter flag mask and tc mask mask_tc_16x8 = _mm_loadl_epi64((__m128i *)(&mask_tc)); mask_flag_q_16x8b = _mm_loadl_epi64((__m128i *)(&mask_flag)); //generating offset 4 const_16x8b = _mm_cmpeq_epi16(tmp_p0_16x8b, tmp_p0_16x8b); // filter flag mask and tc mask mask_tc_16x8 = _mm_shuffle_epi32(mask_tc_16x8, 0x00); mask_flag_q_16x8b = _mm_srai_epi32(mask_flag_q_16x8b, 31); //-tc min_0_16x8b = _mm_sign_epi16(mask_tc_16x8, const_16x8b); //converting const 1 const_16x8b = _mm_srli_epi16(const_16x8b, 15); //filterp mask_flag_p_16x8b = _mm_shuffle_epi32(mask_flag_q_16x8b, 0x00); //converting const 4 const_16x8b = _mm_slli_epi16(const_16x8b, 2); //modified delta with a filter (1 -4 4 -1) available in 16 bit delta_vu0_16x8b = _mm_add_epi16(delta_vu0_16x8b, delta_vu1_16x8b); //filterq flag mask_flag_q_16x8b = _mm_shuffle_epi32(mask_flag_q_16x8b, 0x55); //offset addition delta_vu0_16x8b = _mm_add_epi16(delta_vu0_16x8b, const_16x8b); mask_16x8b = _mm_setzero_si128(); //filter after normalisation delta_vu0_16x8b = _mm_srai_epi16(delta_vu0_16x8b, 3); //converting p0 to 16bit src_p0_16x8b = _mm_unpacklo_epi8(src_p0_16x8b, mask_16x8b); //clipping MAX delta_vu0_16x8b = _mm_min_epi16(delta_vu0_16x8b, mask_tc_16x8); //converting q0 to 16bit src_q0_16x8b = _mm_unpacklo_epi8(src_q0_16x8b, mask_16x8b); //clipping MIN delta_vu0_16x8b = _mm_max_epi16(delta_vu0_16x8b, min_0_16x8b); //masking filter flag delta_vu1_16x8b = _mm_and_si128(delta_vu0_16x8b, mask_flag_q_16x8b); delta_vu0_16x8b = _mm_and_si128(delta_vu0_16x8b, mask_flag_p_16x8b); // q-delta ,p+delta src_q0_16x8b = _mm_sub_epi16(src_q0_16x8b, delta_vu1_16x8b); src_p0_16x8b = _mm_add_epi16(src_p0_16x8b, delta_vu0_16x8b); // p0 and q0 packed src_q0_16x8b = _mm_packus_epi16(src_q0_16x8b, mask_16x8b); src_p0_16x8b = _mm_packus_epi16(src_p0_16x8b, mask_16x8b); _mm_storel_epi64((__m128i *)(pu1_src - src_strd), src_p0_16x8b); _mm_storel_epi64((__m128i *)(pu1_src), src_q0_16x8b); } }