/****************************************************************************** * * Copyright (C) 2018 The Android Open Source Project * * 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. * ***************************************************************************** * Originally developed and contributed by Ittiam Systems Pvt. Ltd, Bangalore */ /** ******************************************************************************* * @file * ihevce_coarse_layer_sad_neon.c * * @brief * Contains intrinsic definitions of functions for computing sad * * @author * Ittiam * * @par List of Functions: * * @remarks * None * ******************************************************************************** */ /*****************************************************************************/ /* File Includes */ /*****************************************************************************/ /* System include files */ #include #include #include #include /* User include files */ #include "ihevc_typedefs.h" #include "itt_video_api.h" #include "ihevc_cmn_utils_neon.h" #include "ihevc_chroma_itrans_recon.h" #include "ihevc_chroma_intra_pred.h" #include "ihevc_debug.h" #include "ihevc_deblk.h" #include "ihevc_defs.h" #include "ihevc_itrans_recon.h" #include "ihevc_intra_pred.h" #include "ihevc_inter_pred.h" #include "ihevc_macros.h" #include "ihevc_mem_fns.h" #include "ihevc_padding.h" #include "ihevc_quant_iquant_ssd.h" #include "ihevc_resi_trans.h" #include "ihevc_sao.h" #include "ihevc_structs.h" #include "ihevc_weighted_pred.h" #include "rc_cntrl_param.h" #include "rc_frame_info_collector.h" #include "rc_look_ahead_params.h" #include "ihevce_api.h" #include "ihevce_defs.h" #include "ihevce_lap_enc_structs.h" #include "ihevce_multi_thrd_structs.h" #include "ihevce_function_selector.h" #include "ihevce_me_common_defs.h" #include "ihevce_enc_structs.h" #include "ihevce_had_satd.h" #include "ihevce_ipe_instr_set_router.h" #include "ihevce_global_tables.h" #include "hme_datatype.h" #include "hme_common_defs.h" #include "hme_interface.h" #include "hme_defs.h" #include "hme_globals.h" #include "ihevce_me_instr_set_router.h" /*****************************************************************************/ /* Function Definitions */ /*****************************************************************************/ void hme_store_4x4_sads_high_speed_neon( hme_search_prms_t *ps_search_prms, layer_ctxt_t *ps_layer_ctxt, range_prms_t *ps_mv_limit, wgt_pred_ctxt_t *ps_wt_inp_prms, S16 *pi2_sads_4x4) { uint8x8_t src2[4]; uint8x16_t src; S32 i, j; /* Input and reference attributes */ U08 *pu1_inp, *pu1_ref; S32 i4_inp_stride, i4_ref_stride, i4_ref_offset; /* The reference is actually an array of ptrs since there are several */ /* reference id. So an array gets passed form calling function */ U08 **ppu1_ref, *pu1_ref_coloc; S32 stepy, stepx, step_shift_x, step_shift_y; S32 mvx, mvy, mv_x_offset, mv_y_offset, mv_x_range, mv_y_range; /* Points to the range limits for mv */ range_prms_t *ps_range_prms = ps_search_prms->aps_mv_range[0]; /* Reference index to be searched */ S32 i4_search_idx = ps_search_prms->i1_ref_idx; pu1_inp = ps_wt_inp_prms->apu1_wt_inp[i4_search_idx]; i4_inp_stride = ps_search_prms->i4_inp_stride; /* Move to the location of the search blk in inp buffer */ pu1_inp += ps_search_prms->i4_cu_x_off; pu1_inp += ps_search_prms->i4_cu_y_off * i4_inp_stride; /*************************************************************************/ /* we use either input of previously encoded pictures as reference */ /* in coarse layer */ /*************************************************************************/ i4_ref_stride = ps_layer_ctxt->i4_inp_stride; ppu1_ref = ps_layer_ctxt->ppu1_list_inp; /* colocated position in reference picture */ i4_ref_offset = (i4_ref_stride * ps_search_prms->i4_y_off) + ps_search_prms->i4_x_off; pu1_ref_coloc = ppu1_ref[i4_search_idx] + i4_ref_offset; stepx = stepy = HME_COARSE_STEP_SIZE_HIGH_SPEED; /*TODO: Calculate Step shift from the #define HME_COARSE_STEP_SIZE_HIGH_SPEED */ step_shift_x = step_shift_y = 2; mv_x_offset = -(ps_mv_limit->i2_min_x >> step_shift_x); mv_y_offset = -(ps_mv_limit->i2_min_y >> step_shift_y); mv_x_range = (-ps_mv_limit->i2_min_x + ps_mv_limit->i2_max_x) >> step_shift_x; mv_y_range = (-ps_mv_limit->i2_min_y + ps_mv_limit->i2_max_y) >> step_shift_y; ASSERT(4 == stepx); /* load input */ { S32 mv_x_sweep = ps_range_prms->i2_max_x - ps_range_prms->i2_min_x; uint32x2_t a[4]; for(i = 0; i < 4; i++) { a[i] = vld1_dup_u32((uint32_t *)pu1_inp); pu1_inp += i4_inp_stride; } src2[0] = vreinterpret_u8_u32(a[0]); src2[1] = vreinterpret_u8_u32(a[1]); src2[2] = vreinterpret_u8_u32(a[2]); src2[3] = vreinterpret_u8_u32(a[3]); if((mv_x_sweep >> step_shift_x) & 1) { uint32x2x2_t l = vtrn_u32(a[0], a[1]); uint32x2x2_t m = vtrn_u32(a[2], a[3]); src = vcombine_u8(vreinterpret_u8_u32(l.val[0]), vreinterpret_u8_u32(m.val[0])); } } /* Run 2loops to sweep over the reference area */ for(mvy = ps_range_prms->i2_min_y; mvy < ps_range_prms->i2_max_y; mvy += stepy) { for(mvx = ps_range_prms->i2_min_x; mvx < ps_range_prms->i2_max_x;) { U16 *pu2_sad = (U16 *)&pi2_sads_4x4 [((mvx >> step_shift_x) + mv_x_offset) + ((mvy >> step_shift_y) + mv_y_offset) * mv_x_range]; pu1_ref = pu1_ref_coloc + mvx + (mvy * i4_ref_stride); if((mvx + (stepx * 4)) <= ps_range_prms->i2_max_x) // 16x4 { uint16x8_t abs_01 = vdupq_n_u16(0); uint16x8_t abs_23 = vdupq_n_u16(0); uint16x4_t tmp_a0, tmp_a1; for(j = 0; j < 4; j++) { uint8x16_t ref = vld1q_u8(pu1_ref); abs_01 = vabal_u8(abs_01, src2[j], vget_low_u8(ref)); abs_23 = vabal_u8(abs_23, src2[j], vget_high_u8(ref)); pu1_ref += i4_ref_stride; } tmp_a0 = vpadd_u16(vget_low_u16(abs_01), vget_high_u16(abs_01)); tmp_a1 = vpadd_u16(vget_low_u16(abs_23), vget_high_u16(abs_23)); abs_01 = vcombine_u16(tmp_a0, tmp_a1); tmp_a0 = vpadd_u16(vget_low_u16(abs_01), vget_high_u16(abs_01)); vst1_u16(pu2_sad, tmp_a0); mvx += stepx * 4; } else if((mvx + (stepx * 2)) <= ps_range_prms->i2_max_x) // 8x4 { uint16x8_t abs_01 = vdupq_n_u16(0); uint16x4_t tmp_a; uint32x2_t tmp_b; for(j = 0; j < 4; j++) { uint8x8_t ref = vld1_u8(pu1_ref); abs_01 = vabal_u8(abs_01, src2[j], ref); pu1_ref += i4_ref_stride; } tmp_a = vpadd_u16(vget_low_u16(abs_01), vget_high_u16(abs_01)); tmp_b = vpaddl_u16(tmp_a); pu2_sad[0] = vget_lane_u32(tmp_b, 0); pu2_sad[1] = vget_lane_u32(tmp_b, 1); mvx += stepx * 2; } else if((mvx + stepx) <= ps_range_prms->i2_max_x) // 4x4 { const uint8x16_t ref = load_unaligned_u8q(pu1_ref, i4_ref_stride); uint16x8_t abs = vabdl_u8(vget_low_u8(src), vget_low_u8(ref)); uint32x4_t b; uint64x2_t c; abs = vabal_u8(abs, vget_high_u8(src), vget_high_u8(ref)); b = vpaddlq_u16(abs); c = vpaddlq_u32(b); *pu2_sad = vget_lane_u32( vadd_u32( vreinterpret_u32_u64(vget_low_u64(c)), vreinterpret_u32_u64(vget_high_u64(c))), 0); mvx += stepx; } } } } void hme_store_4x4_sads_high_quality_neon( hme_search_prms_t *ps_search_prms, layer_ctxt_t *ps_layer_ctxt, range_prms_t *ps_mv_limit, wgt_pred_ctxt_t *ps_wt_inp_prms, S16 *pi2_sads_4x4) { uint8x8_t src2[4]; uint8x16_t src; S32 i, j; /* Input and reference attributes */ U08 *pu1_inp, *pu1_ref; S32 i4_inp_stride, i4_ref_stride, i4_ref_offset; /* The reference is actually an array of ptrs since there are several */ /* reference id. So an array gets passed form calling function */ U08 **ppu1_ref, *pu1_ref_coloc; S32 stepy, stepx, step_shift_x, step_shift_y; S32 mvx, mvy, mv_x_offset, mv_y_offset, mv_x_range, mv_y_range; /* Points to the range limits for mv */ range_prms_t *ps_range_prms = ps_search_prms->aps_mv_range[0]; /* Reference index to be searched */ S32 i4_search_idx = ps_search_prms->i1_ref_idx; pu1_inp = ps_wt_inp_prms->apu1_wt_inp[i4_search_idx]; i4_inp_stride = ps_search_prms->i4_inp_stride; /* Move to the location of the search blk in inp buffer */ pu1_inp += ps_search_prms->i4_cu_x_off; pu1_inp += ps_search_prms->i4_cu_y_off * i4_inp_stride; /*************************************************************************/ /* we use either input of previously encoded pictures as reference */ /* in coarse layer */ /*************************************************************************/ i4_ref_stride = ps_layer_ctxt->i4_inp_stride; ppu1_ref = ps_layer_ctxt->ppu1_list_inp; /* colocated position in reference picture */ i4_ref_offset = (i4_ref_stride * ps_search_prms->i4_y_off) + ps_search_prms->i4_x_off; pu1_ref_coloc = ppu1_ref[i4_search_idx] + i4_ref_offset; stepx = stepy = HME_COARSE_STEP_SIZE_HIGH_QUALITY; /*TODO: Calculate Step shift from the #define HME_COARSE_STEP_SIZE_HIGH_QUALITY */ step_shift_x = step_shift_y = 1; mv_x_offset = -(ps_mv_limit->i2_min_x >> step_shift_x); mv_y_offset = -(ps_mv_limit->i2_min_y >> step_shift_y); mv_x_range = (-ps_mv_limit->i2_min_x + ps_mv_limit->i2_max_x) >> step_shift_x; mv_y_range = (-ps_mv_limit->i2_min_y + ps_mv_limit->i2_max_y) >> step_shift_y; /* load input */ { S32 mv_x_sweep = ps_range_prms->i2_max_x - ps_range_prms->i2_min_x; uint32x2_t a[4]; for(i = 0; i < 4; i++) { a[i] = vld1_dup_u32((uint32_t *)pu1_inp); pu1_inp += i4_inp_stride; } src2[0] = vreinterpret_u8_u32(a[0]); src2[1] = vreinterpret_u8_u32(a[1]); src2[2] = vreinterpret_u8_u32(a[2]); src2[3] = vreinterpret_u8_u32(a[3]); if((mv_x_sweep >> 2) & 1) { uint32x2x2_t l = vtrn_u32(a[0], a[1]); uint32x2x2_t m = vtrn_u32(a[2], a[3]); src = vcombine_u8(vreinterpret_u8_u32(l.val[0]), vreinterpret_u8_u32(m.val[0])); } } /* Run 2loops to sweep over the reference area */ for(mvy = ps_range_prms->i2_min_y; mvy < ps_range_prms->i2_max_y; mvy += stepy) { for(mvx = ps_range_prms->i2_min_x; mvx < ps_range_prms->i2_max_x;) { U16 *pu2_sad = (U16 *)&pi2_sads_4x4 [((mvx >> step_shift_x) + mv_x_offset) + ((mvy >> step_shift_y) + mv_y_offset) * mv_x_range]; pu1_ref = pu1_ref_coloc + mvx + (mvy * i4_ref_stride); if((mvx + (stepx * 8)) <= ps_range_prms->i2_max_x) // 16x4 { uint16x8_t abs_a_01 = vdupq_n_u16(0); uint16x8_t abs_a_23 = vdupq_n_u16(0); uint16x8_t abs_b_01 = vdupq_n_u16(0); uint16x8_t abs_b_23 = vdupq_n_u16(0); uint16x4_t tmp_b0, tmp_b1; uint16x4x2_t tmp_a; for(j = 0; j < 4; j++) { uint8x16_t ref_a = vld1q_u8(pu1_ref); uint8x16_t ref_b = vld1q_u8(pu1_ref + 2); abs_a_01 = vabal_u8(abs_a_01, src2[j], vget_low_u8(ref_a)); abs_a_23 = vabal_u8(abs_a_23, src2[j], vget_high_u8(ref_a)); abs_b_01 = vabal_u8(abs_b_01, src2[j], vget_low_u8(ref_b)); abs_b_23 = vabal_u8(abs_b_23, src2[j], vget_high_u8(ref_b)); pu1_ref += i4_ref_stride; } tmp_a.val[0] = vpadd_u16(vget_low_u16(abs_a_01), vget_high_u16(abs_a_01)); tmp_a.val[1] = vpadd_u16(vget_low_u16(abs_a_23), vget_high_u16(abs_a_23)); abs_a_01 = vcombine_u16(tmp_a.val[0], tmp_a.val[1]); tmp_a.val[0] = vpadd_u16(vget_low_u16(abs_a_01), vget_high_u16(abs_a_01)); tmp_b0 = vpadd_u16(vget_low_u16(abs_b_01), vget_high_u16(abs_b_01)); tmp_b1 = vpadd_u16(vget_low_u16(abs_b_23), vget_high_u16(abs_b_23)); abs_b_01 = vcombine_u16(tmp_b0, tmp_b1); tmp_a.val[1] = vpadd_u16(vget_low_u16(abs_b_01), vget_high_u16(abs_b_01)); vst2_u16(pu2_sad, tmp_a); mvx += stepx * 8; } else if((mvx + (stepx * 4)) <= ps_range_prms->i2_max_x) // 8x4 { uint16x8_t abs_a_01 = vdupq_n_u16(0); uint16x8_t abs_b_01 = vdupq_n_u16(0); uint16x4_t tmp_a, tmp_b; for(j = 0; j < 4; j++) { uint8x8_t ref_a = vld1_u8(pu1_ref); uint8x8_t ref_b = vld1_u8(pu1_ref + 2); abs_a_01 = vabal_u8(abs_a_01, src2[j], ref_a); abs_b_01 = vabal_u8(abs_b_01, src2[j], ref_b); pu1_ref += i4_ref_stride; } tmp_a = vpadd_u16(vget_low_u16(abs_a_01), vget_high_u16(abs_a_01)); tmp_b = vpadd_u16(vget_low_u16(abs_b_01), vget_high_u16(abs_b_01)); tmp_a = vpadd_u16(tmp_a, tmp_b); pu2_sad[0] = vget_lane_u16(tmp_a, 0); pu2_sad[1] = vget_lane_u16(tmp_a, 2); pu2_sad[2] = vget_lane_u16(tmp_a, 1); pu2_sad[3] = vget_lane_u16(tmp_a, 3); mvx += stepx * 4; } else if((mvx + stepx * 2) <= ps_range_prms->i2_max_x) // 4x4 { uint8x16_t ref = load_unaligned_u8q(pu1_ref, i4_ref_stride); uint16x8_t abs = vabdl_u8(vget_low_u8(src), vget_low_u8(ref)); uint32x4_t b; uint64x2_t c; abs = vabal_u8(abs, vget_high_u8(src), vget_high_u8(ref)); b = vpaddlq_u16(abs); c = vpaddlq_u32(b); *pu2_sad = vget_lane_u32( vadd_u32( vreinterpret_u32_u64(vget_low_u64(c)), vreinterpret_u32_u64(vget_high_u64(c))), 0); ref = load_unaligned_u8q(pu1_ref + 2, i4_ref_stride); abs = vabdl_u8(vget_low_u8(src), vget_low_u8(ref)); abs = vabal_u8(abs, vget_high_u8(src), vget_high_u8(ref)); b = vpaddlq_u16(abs); c = vpaddlq_u32(b); pu2_sad[1] = vget_lane_u32( vadd_u32( vreinterpret_u32_u64(vget_low_u64(c)), vreinterpret_u32_u64(vget_high_u64(c))), 0); mvx += stepx * 2; } else { assert(0); } } } } #define BEST_COST(i) \ if(sad_array[0][i] < min_cost_4x8) \ { \ best_mv_x_4x8 = mvx + i * stepx; \ best_mv_y_4x8 = mvy; \ min_cost_4x8 = sad_array[0][i]; \ } \ if(sad_array[1][i] < min_cost_8x4) \ { \ best_mv_x_8x4 = mvx + i * stepx; \ best_mv_y_8x4 = mvy; \ min_cost_8x4 = sad_array[1][i]; \ } void hme_combine_4x4_sads_and_compute_cost_high_speed_neon( S08 i1_ref_idx, range_prms_t *ps_mv_range, range_prms_t *ps_mv_limit, hme_mv_t *ps_best_mv_4x8, hme_mv_t *ps_best_mv_8x4, pred_ctxt_t *ps_pred_ctxt, PF_MV_COST_FXN pf_mv_cost_compute, S16 *pi2_sads_4x4_current, S16 *pi2_sads_4x4_east, S16 *pi2_sads_4x4_south) { S32 best_mv_y_4x8, best_mv_x_4x8, best_mv_y_8x4, best_mv_x_8x4; S32 stepy = HME_COARSE_STEP_SIZE_HIGH_SPEED; S32 stepx = HME_COARSE_STEP_SIZE_HIGH_SPEED; /*TODO: Calculate Step shift from the #define HME_COARSE_STEP_SIZE_HIGH_SPEED */ S32 step_shift_x = 2; S32 step_shift_y = 2; S32 mvx, mvy, mv_x_offset, mv_y_offset, mv_x_range, mv_y_range; S32 lambda = ps_pred_ctxt->lambda; S32 lambda_q_shift = ps_pred_ctxt->lambda_q_shift; S32 rnd = 1 << (lambda_q_shift - 1); S32 min_cost_4x8 = MAX_32BIT_VAL; S32 min_cost_8x4 = MAX_32BIT_VAL; S32 i; const uint16x8_t v_ref_idx = vdupq_n_u16(i1_ref_idx); const uint32x4_t v_lambda = vdupq_n_u32(lambda); const uint32x4_t v_rnd_factor = vdupq_n_u32(rnd); const int32x4_t v_lambda_q_shift = vdupq_n_s32(-lambda_q_shift); mv_x_offset = (-ps_mv_limit->i2_min_x >> step_shift_x); mv_y_offset = (-ps_mv_limit->i2_min_y >> step_shift_y); mv_x_range = (-ps_mv_limit->i2_min_x + ps_mv_limit->i2_max_x) >> step_shift_x; mv_y_range = (-ps_mv_limit->i2_min_y + ps_mv_limit->i2_max_y) >> step_shift_y; ASSERT(MAX_MVX_SUPPORTED_IN_COARSE_LAYER >= ABS(ps_mv_range->i2_max_x)); ASSERT(MAX_MVY_SUPPORTED_IN_COARSE_LAYER >= ABS(ps_mv_range->i2_max_y)); /* Run 2loops to sweep over the reference area */ for(mvy = ps_mv_range->i2_min_y; mvy < ps_mv_range->i2_max_y; mvy += stepy) { /* LUT: (2 * hme_get_range(mv_y) - 1) + ((!mv_y) ? 0 : 1) */ uint16x8_t mvy_wt = vld1q_u16((U16 *)&gi2_mvy_range[ABS(mvy)][0]); /* mvy wt + ref_idx */ mvy_wt = vaddq_u16(mvy_wt, v_ref_idx); for(mvx = ps_mv_range->i2_min_x; mvx < ps_mv_range->i2_max_x;) { S32 sad_pos = ((mvx >> step_shift_x) + mv_x_offset) + ((mvy >> step_shift_y) + mv_y_offset) * mv_x_range; if(mvx + (8 * stepx) <= ps_mv_range->i2_max_x) // 8x4 { uint16x8_t curr = vld1q_u16((U16 *)pi2_sads_4x4_current + sad_pos); uint16x8_t south = vld1q_u16((U16 *)pi2_sads_4x4_south + sad_pos); uint16x8_t east = vld1q_u16((U16 *)pi2_sads_4x4_east + sad_pos); uint16x8_t sad_4x8 = vaddq_u16(curr, south); uint16x8_t sad_8x4 = vaddq_u16(curr, east); /* LUT: (2 * hme_get_range(mv_x) - 1) + ((!mv_x) ? 0 : 1) */ uint16x8_t mv_wt = vld1q_u16((U16 *)&gi2_mvx_range[mvx + MAX_MVX_SUPPORTED_IN_COARSE_LAYER][0]); uint32x4_t total_cost_0, total_cost_1; uint16x8_t total_cost; U16 sad_array[2][8]; /* mv weight + ref_idx */ mv_wt = vaddq_u16(mv_wt, mvy_wt); total_cost_0 = vmulq_u32(v_lambda, vmovl_u16(vget_low_u16(mv_wt))); total_cost_1 = vmulq_u32(v_lambda, vmovl_u16(vget_high_u16(mv_wt))); total_cost_0 = vaddq_u32(total_cost_0, v_rnd_factor); total_cost_1 = vaddq_u32(total_cost_1, v_rnd_factor); total_cost_0 = vshlq_u32(total_cost_0, v_lambda_q_shift); total_cost_1 = vshlq_u32(total_cost_1, v_lambda_q_shift); total_cost = vcombine_u16(vmovn_u32(total_cost_0), vmovn_u32(total_cost_1)); sad_4x8 = vaddq_u16(total_cost, sad_4x8); sad_8x4 = vaddq_u16(total_cost, sad_8x4); vst1q_u16(sad_array[0], sad_4x8); vst1q_u16(sad_array[1], sad_8x4); for(i = 0; i < 8; i++) { BEST_COST(i); } mvx += stepx * 8; } else if(mvx + (4 * stepx) <= ps_mv_range->i2_max_x) // 4x4 { uint16x4_t curr = vld1_u16((U16 *)pi2_sads_4x4_current + sad_pos); uint16x4_t south = vld1_u16((U16 *)pi2_sads_4x4_south + sad_pos); uint16x4_t east = vld1_u16((U16 *)pi2_sads_4x4_east + sad_pos); uint16x4_t sad_4x8 = vadd_u16(curr, south); uint16x4_t sad_8x4 = vadd_u16(curr, east); /* LUT: (2 * hme_get_range(mv_x) - 1) + ((!mv_x) ? 0 : 1) */ uint16x4_t mv_wt = vld1_u16((U16 *)&gi2_mvx_range[mvx + MAX_MVX_SUPPORTED_IN_COARSE_LAYER][0]); uint32x4_t total_cost; U16 sad_array[2][4]; /* mv weight + ref_idx */ mv_wt = vadd_u16(mv_wt, vget_low_u16(mvy_wt)); total_cost = vmulq_u32(v_lambda, vmovl_u16(mv_wt)); total_cost = vaddq_u32(total_cost, v_rnd_factor); total_cost = vshlq_u32(total_cost, v_lambda_q_shift); sad_4x8 = vadd_u16(vmovn_u32(total_cost), sad_4x8); sad_8x4 = vadd_u16(vmovn_u32(total_cost), sad_8x4); vst1_u16(sad_array[0], sad_4x8); vst1_u16(sad_array[1], sad_8x4); for(i = 0; i < 4; i++) { BEST_COST(i); } mvx += stepx * 4; } else { S16 sad_array[2][1]; S32 mv_cost; /* Get SAD by adding SAD for current and neighbour S */ sad_array[0][0] = pi2_sads_4x4_current[sad_pos] + pi2_sads_4x4_south[sad_pos]; sad_array[1][0] = pi2_sads_4x4_current[sad_pos] + pi2_sads_4x4_east[sad_pos]; mv_cost = gi2_mvy_range[ABS(mvy)][0] + gi2_mvx_range[mvx + MAX_MVX_SUPPORTED_IN_COARSE_LAYER][0] + i1_ref_idx; mv_cost = (mv_cost * lambda + rnd) >> lambda_q_shift; sad_array[0][0] += mv_cost; sad_array[1][0] += mv_cost; BEST_COST(0); mvx += stepx; } } } ps_best_mv_4x8->i2_mv_x = best_mv_x_4x8; ps_best_mv_4x8->i2_mv_y = best_mv_y_4x8; ps_best_mv_8x4->i2_mv_x = best_mv_x_8x4; ps_best_mv_8x4->i2_mv_y = best_mv_y_8x4; } void hme_combine_4x4_sads_and_compute_cost_high_quality_neon( S08 i1_ref_idx, range_prms_t *ps_mv_range, range_prms_t *ps_mv_limit, hme_mv_t *ps_best_mv_4x8, hme_mv_t *ps_best_mv_8x4, pred_ctxt_t *ps_pred_ctxt, PF_MV_COST_FXN pf_mv_cost_compute, S16 *pi2_sads_4x4_current, S16 *pi2_sads_4x4_east, S16 *pi2_sads_4x4_south) { S32 best_mv_y_4x8, best_mv_x_4x8, best_mv_y_8x4, best_mv_x_8x4; S32 stepy = HME_COARSE_STEP_SIZE_HIGH_QUALITY; S32 stepx = HME_COARSE_STEP_SIZE_HIGH_QUALITY; /*TODO: Calculate Step shift from the #define HME_COARSE_STEP_SIZE_HIGH_SPEED */ S32 step_shift_x = 1; S32 step_shift_y = 1; S32 mvx, mvy, mv_x_offset, mv_y_offset, mv_x_range, mv_y_range; S32 lambda = ps_pred_ctxt->lambda; S32 lambda_q_shift = ps_pred_ctxt->lambda_q_shift; S32 rnd = 1 << (lambda_q_shift - 1); S32 min_cost_4x8 = MAX_32BIT_VAL; S32 min_cost_8x4 = MAX_32BIT_VAL; S32 i; const uint16x8_t v_ref_idx = vdupq_n_u16(i1_ref_idx); const uint32x4_t v_lambda = vdupq_n_u32(lambda); const uint32x4_t v_rnd_factor = vdupq_n_u32(rnd); const int32x4_t v_lambda_q_shift = vdupq_n_s32(-lambda_q_shift); mv_x_offset = (-ps_mv_limit->i2_min_x >> step_shift_x); mv_y_offset = (-ps_mv_limit->i2_min_y >> step_shift_y); mv_x_range = (-ps_mv_limit->i2_min_x + ps_mv_limit->i2_max_x) >> step_shift_x; mv_y_range = (-ps_mv_limit->i2_min_y + ps_mv_limit->i2_max_y) >> step_shift_y; ASSERT(MAX_MVX_SUPPORTED_IN_COARSE_LAYER >= ABS(ps_mv_range->i2_max_x)); ASSERT(MAX_MVY_SUPPORTED_IN_COARSE_LAYER >= ABS(ps_mv_range->i2_max_y)); /* Run 2loops to sweep over the reference area */ for(mvy = ps_mv_range->i2_min_y; mvy < ps_mv_range->i2_max_y; mvy += stepy) { /* LUT: (2 * hme_get_range(mv_y) - 1) + ((!mv_y) ? 0 : 1) */ uint16x8_t mvy_wt = vld1q_u16((U16 *)&gi2_mvy_range[ABS(mvy)][0]); /* mvy wt + ref_idx */ mvy_wt = vaddq_u16(mvy_wt, v_ref_idx); for(mvx = ps_mv_range->i2_min_x; mvx < ps_mv_range->i2_max_x;) { S32 sad_pos = ((mvx >> step_shift_x) + mv_x_offset) + ((mvy >> step_shift_y) + mv_y_offset) * mv_x_range; if(mvx + (8 * stepx) <= ps_mv_range->i2_max_x) // 8x4 { uint16x8_t curr = vld1q_u16((U16 *)pi2_sads_4x4_current + sad_pos); uint16x8_t south = vld1q_u16((U16 *)pi2_sads_4x4_south + sad_pos); uint16x8_t east = vld1q_u16((U16 *)pi2_sads_4x4_east + sad_pos); uint16x8_t sad_4x8 = vaddq_u16(curr, south); uint16x8_t sad_8x4 = vaddq_u16(curr, east); /* LUT: (2 * hme_get_range(mv_x) - 1) + ((!mv_x) ? 0 : 1) */ uint16x8_t mv_wt = vld1q_u16( (U16 *)&gi2_mvx_range_high_quality[mvx + MAX_MVX_SUPPORTED_IN_COARSE_LAYER][0]); uint32x4_t total_cost_0, total_cost_1; uint16x8_t total_cost; U16 sad_array[2][8]; /* mv weight + ref_idx */ mv_wt = vaddq_u16(mv_wt, mvy_wt); total_cost_0 = vmulq_u32(v_lambda, vmovl_u16(vget_low_u16(mv_wt))); total_cost_1 = vmulq_u32(v_lambda, vmovl_u16(vget_high_u16(mv_wt))); total_cost_0 = vaddq_u32(total_cost_0, v_rnd_factor); total_cost_1 = vaddq_u32(total_cost_1, v_rnd_factor); total_cost_0 = vshlq_u32(total_cost_0, v_lambda_q_shift); total_cost_1 = vshlq_u32(total_cost_1, v_lambda_q_shift); total_cost = vcombine_u16(vmovn_u32(total_cost_0), vmovn_u32(total_cost_1)); sad_4x8 = vaddq_u16(total_cost, sad_4x8); sad_8x4 = vaddq_u16(total_cost, sad_8x4); vst1q_u16(sad_array[0], sad_4x8); vst1q_u16(sad_array[1], sad_8x4); for(i = 0; i < 8; i++) { BEST_COST(i); } mvx += stepx * 8; } else if(mvx + (4 * stepx) <= ps_mv_range->i2_max_x) // 4x4 { uint16x4_t curr = vld1_u16((U16 *)pi2_sads_4x4_current + sad_pos); uint16x4_t south = vld1_u16((U16 *)pi2_sads_4x4_south + sad_pos); uint16x4_t east = vld1_u16((U16 *)pi2_sads_4x4_east + sad_pos); uint16x4_t sad_4x8 = vadd_u16(curr, south); uint16x4_t sad_8x4 = vadd_u16(curr, east); /* LUT: (2 * hme_get_range(mv_x) - 1) + ((!mv_x) ? 0 : 1) */ uint16x4_t mv_wt = vld1_u16( (U16 *)&gi2_mvx_range_high_quality[mvx + MAX_MVX_SUPPORTED_IN_COARSE_LAYER][0]); uint32x4_t total_cost; U16 sad_array[2][4]; /* mv weight + ref_idx */ mv_wt = vadd_u16(mv_wt, vget_low_u16(mvy_wt)); total_cost = vmulq_u32(v_lambda, vmovl_u16(mv_wt)); total_cost = vaddq_u32(total_cost, v_rnd_factor); total_cost = vshlq_u32(total_cost, v_lambda_q_shift); sad_4x8 = vadd_u16(vmovn_u32(total_cost), sad_4x8); sad_8x4 = vadd_u16(vmovn_u32(total_cost), sad_8x4); vst1_u16(sad_array[0], sad_4x8); vst1_u16(sad_array[1], sad_8x4); for(i = 0; i < 4; i++) { BEST_COST(i); } mvx += stepx * 4; } else { S16 sad_array[2][1]; S32 mv_cost; /* Get SAD by adding SAD for current and neighbour S */ sad_array[0][0] = pi2_sads_4x4_current[sad_pos] + pi2_sads_4x4_south[sad_pos]; sad_array[1][0] = pi2_sads_4x4_current[sad_pos] + pi2_sads_4x4_east[sad_pos]; mv_cost = gi2_mvy_range[ABS(mvy)][0] + gi2_mvx_range_high_quality[mvx + MAX_MVX_SUPPORTED_IN_COARSE_LAYER][0] + i1_ref_idx; mv_cost = (mv_cost * lambda + rnd) >> lambda_q_shift; sad_array[0][0] += mv_cost; sad_array[1][0] += mv_cost; BEST_COST(0); mvx += stepx; } } } ps_best_mv_4x8->i2_mv_x = best_mv_x_4x8; ps_best_mv_4x8->i2_mv_y = best_mv_y_4x8; ps_best_mv_8x4->i2_mv_x = best_mv_x_8x4; ps_best_mv_8x4->i2_mv_y = best_mv_y_8x4; }