@/****************************************************************************** @ * @ * Copyright (C) 2015 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 @* ih264_weighted_bi_pred_a9q.s @* @* @brief @* Contains function definitions for weighted biprediction. @* @* @author @* Kaushik Senthoor R @* @* @par List of Functions: @* @* - ih264_weighted_bi_pred_luma_a9q() @* - ih264_weighted_bi_pred_chroma_a9q() @* @* @remarks @* None @* @******************************************************************************* @* @******************************************************************************* @* @function @* ih264_weighted_bi_pred_luma_a9q() @* @* @brief @* This routine performs the weighted biprediction as described in sec @* 8.4.2.3.2 titled "Weighted sample prediction process" for luma. @* @* @par Description: @* This function gets two ht x wd blocks, calculates the weighted samples, @* rounds off, adds offset and stores it in the destination block. @* @* @param[in] pu1_src1 @* UWORD8 Pointer to the buffer containing the input block 1. @* @* @param[in] pu1_src2 @* UWORD8 Pointer to the buffer containing the input block 2. @* @* @param[out] pu1_dst @* UWORD8 pointer to the destination where the output block is stored. @* @* @param[in] src_strd1 @* Stride of the input buffer 1 @* @* @param[in] src_strd2 @* Stride of the input buffer 2 @* @* @param[in] dst_strd @* Stride of the destination buffer @* @* @param[in] log_wd @* number of bits to be rounded off @* @* @param[in] wt1 @* weight for the weighted prediction @* @* @param[in] wt2 @* weight for the weighted prediction @* @* @param[in] ofst1 @* offset 1 used after rounding off @* @* @param[in] ofst2 @* offset 2 used after rounding off @* @* @param[in] ht @* integer height of the array @* @* @param[in] wd @* integer width of the array @* @* @returns @* None @* @* @remarks @* (ht,wd) can be (4,4), (4,8), (8,4), (8,8), (8,16), (16,8) or (16,16). @* @******************************************************************************* @* @void ih264_weighted_bi_pred_luma_a9q(UWORD8 *pu1_src1, @ UWORD8 *pu1_src2, @ UWORD8 *pu1_dst, @ WORD32 src_strd1, @ WORD32 src_strd2, @ WORD32 dst_strd, @ WORD32 log_wd, @ WORD32 wt1, @ WORD32 wt2, @ WORD32 ofst1, @ WORD32 ofst2, @ WORD32 ht, @ WORD32 wd) @ @**************Variables Vs Registers***************************************** @ r0 => pu1_src1 @ r1 => pu1_src2 @ r2 => pu1_dst @ r3 => src_strd1 @ [sp] => src_strd2 (r4) @ [sp+4] => dst_strd (r5) @ [sp+8] => log_wd (r6) @ [sp+12] => wt1 (r7) @ [sp+16] => wt2 (r8) @ [sp+20] => ofst1 (r9) @ [sp+24] => ofst2 (r10) @ [sp+28] => ht (r11) @ [sp+32] => wd (r12) @ .text .p2align 2 .global ih264_weighted_bi_pred_luma_a9q ih264_weighted_bi_pred_luma_a9q: stmfd sp!, {r4-r12, r14} @stack stores the values of the arguments ldr r6, [sp, #48] @Load log_wd in r6 ldr r7, [sp, #52] @Load wt1 in r7 ldr r8, [sp, #56] @Load wt2 in r8 ldr r9, [sp, #60] @Load ofst1 in r9 add r6, r6, #1 @r6 = log_wd + 1 sxtb r7, r7 @sign-extend 16-bit wt1 to 32-bit ldr r4, [sp, #40] @Load src_strd2 in r4 ldr r5, [sp, #44] @Load dst_strd in r5 sxtb r9, r9 @sign-extend 8-bit ofst1 to 32-bit neg r10, r6 @r10 = -(log_wd + 1) ldr r11, [sp, #68] @Load ht in r11 ldr r12, [sp, #72] @Load wd in r12 vdup.16 q0, r10 @Q0 = -(log_wd + 1) (32-bit) add r9, r9, #1 @r9 = ofst1 + 1 ldr r10, [sp, #64] @Load ofst2 in r10 sxtb r8, r8 @sign-extend 16-bit wt2 to 32-bit cmp r12, #16 @check if wd is 16 vpush {d8-d15} sxtb r10, r10 @sign-extend 8-bit ofst2 to 32-bit add r9, r9, r10 @r9 = ofst1 + ofst2 + 1 vmov d2, r7, r8 @D2 = {wt1(32-bit), wt2(32-bit)} asr r9, r9, #1 @r9 = ofst = (ofst1 + ofst2 + 1) >> 1 vdup.8 d3, r9 @D3 = ofst (8-bit) beq loop_16 @branch if wd is 16 cmp r12, #8 @check if wd is 8 beq loop_8 @branch if wd is 8 loop_4: @each iteration processes four rows vld1.32 d4[0], [r0], r3 @load row 1 in source 1 vld1.32 d4[1], [r0], r3 @load row 2 in source 1 vld1.32 d6[0], [r1], r4 @load row 1 in source 2 vld1.32 d6[1], [r1], r4 @load row 2 in source 2 vmovl.u8 q2, d4 @converting rows 1,2 in source 1 to 16-bit vld1.32 d8[0], [r0], r3 @load row 3 in source 1 vld1.32 d8[1], [r0], r3 @load row 4 in source 1 vmovl.u8 q3, d6 @converting rows 1,2 in source 2 to 16-bit vld1.32 d10[0], [r1], r4 @load row 3 in source 2 vld1.32 d10[1], [r1], r4 @load row 4 in source 2 vmovl.u8 q4, d8 @converting rows 3,4 in source 1 to 16-bit vmovl.u8 q5, d10 @converting rows 3,4 in source 2 to 16-bit vmul.s16 q2, q2, d2[0] @weight 1 mult. for rows 1,2 vmla.s16 q2, q3, d2[2] @weight 2 mult. for rows 1,2 vmul.s16 q4, q4, d2[0] @weight 1 mult. for rows 3,4 vmla.s16 q4, q5, d2[2] @weight 2 mult. for rows 3,4 subs r11, r11, #4 @decrement ht by 4 vrshl.s16 q2, q2, q0 @rounds off the weighted samples from rows 1,2 vrshl.s16 q4, q4, q0 @rounds off the weighted samples from rows 3,4 vaddw.s8 q2, q2, d3 @adding offset for rows 1,2 vaddw.s8 q4, q4, d3 @adding offset for rows 3,4 vqmovun.s16 d4, q2 @saturating rows 1,2 to unsigned 8-bit vqmovun.s16 d8, q4 @saturating rows 3,4 to unsigned 8-bit vst1.32 d4[0], [r2], r5 @store row 1 in destination vst1.32 d4[1], [r2], r5 @store row 2 in destination vst1.32 d8[0], [r2], r5 @store row 3 in destination vst1.32 d8[1], [r2], r5 @store row 4 in destination bgt loop_4 @if greater than 0 repeat the loop again b end_loops loop_8: @each iteration processes four rows vld1.8 d4, [r0], r3 @load row 1 in source 1 vld1.8 d6, [r1], r4 @load row 1 in source 2 vld1.8 d8, [r0], r3 @load row 2 in source 1 vld1.8 d10, [r1], r4 @load row 2 in source 2 vmovl.u8 q2, d4 @converting row 1 in source 1 to 16-bit vld1.8 d12, [r0], r3 @load row 3 in source 1 vld1.8 d14, [r1], r4 @load row 3 in source 2 vmovl.u8 q3, d6 @converting row 1 in source 2 to 16-bit vld1.8 d16, [r0], r3 @load row 4 in source 1 vld1.8 d18, [r1], r4 @load row 4 in source 2 vmovl.u8 q4, d8 @converting row 2 in source 1 to 16-bit vmovl.u8 q5, d10 @converting row 2 in source 2 to 16-bit vmul.s16 q2, q2, d2[0] @weight 1 mult. for row 1 vmla.s16 q2, q3, d2[2] @weight 2 mult. for row 1 vmovl.u8 q6, d12 @converting row 3 in source 1 to 16-bit vmovl.u8 q7, d14 @converting row 3 in source 2 to 16-bit vmul.s16 q4, q4, d2[0] @weight 1 mult. for row 2 vmla.s16 q4, q5, d2[2] @weight 2 mult. for row 2 vmovl.u8 q8, d16 @converting row 4 in source 1 to 16-bit vmovl.u8 q9, d18 @converting row 4 in source 2 to 16-bit vmul.s16 q6, q6, d2[0] @weight 1 mult. for row 3 vmla.s16 q6, q7, d2[2] @weight 2 mult. for row 3 vmul.s16 q8, q8, d2[0] @weight 1 mult. for row 4 vmla.s16 q8, q9, d2[2] @weight 2 mult. for row 4 vrshl.s16 q2, q2, q0 @rounds off the weighted samples from row 1 vrshl.s16 q4, q4, q0 @rounds off the weighted samples from row 2 vrshl.s16 q6, q6, q0 @rounds off the weighted samples from row 3 vaddw.s8 q2, q2, d3 @adding offset for row 1 vrshl.s16 q8, q8, q0 @rounds off the weighted samples from row 4 vaddw.s8 q4, q4, d3 @adding offset for row 2 vaddw.s8 q6, q6, d3 @adding offset for row 3 vqmovun.s16 d4, q2 @saturating row 1 to unsigned 8-bit vaddw.s8 q8, q8, d3 @adding offset for row 4 vqmovun.s16 d8, q4 @saturating row 2 to unsigned 8-bit vqmovun.s16 d12, q6 @saturating row 3 to unsigned 8-bit vqmovun.s16 d16, q8 @saturating row 4 to unsigned 8-bit vst1.8 d4, [r2], r5 @store row 1 in destination vst1.8 d8, [r2], r5 @store row 2 in destination subs r11, r11, #4 @decrement ht by 4 vst1.8 d12, [r2], r5 @store row 3 in destination vst1.8 d16, [r2], r5 @store row 4 in destination bgt loop_8 @if greater than 0 repeat the loop again b end_loops loop_16: @each iteration processes two rows vld1.8 {q2}, [r0], r3 @load row 1 in source 1 vld1.8 {q3}, [r1], r4 @load row 1 in source 2 vld1.8 {q4}, [r0], r3 @load row 2 in source 1 vld1.8 {q5}, [r1], r4 @load row 2 in source 2 vmovl.u8 q10, d4 @converting row 1L in source 1 to 16-bit vld1.8 {q6}, [r0], r3 @load row 3 in source 1 vld1.8 {q7}, [r1], r4 @load row 3 in source 2 vmovl.u8 q11, d6 @converting row 1L in source 2 to 16-bit vld1.8 {q8}, [r0], r3 @load row 4 in source 1 vld1.8 {q9}, [r1], r4 @load row 4 in source 2 vmovl.u8 q2, d5 @converting row 1H in source 1 to 16-bit vmovl.u8 q3, d7 @converting row 1H in source 2 to 16-bit vmul.s16 q10, q10, d2[0] @weight 1 mult. for row 1L vmla.s16 q10, q11, d2[2] @weight 2 mult. for row 1L vmovl.u8 q12, d8 @converting row 2L in source 1 to 16-bit vmovl.u8 q13, d10 @converting row 2L in source 2 to 16-bit vmul.s16 q2, q2, d2[0] @weight 1 mult. for row 1H vmla.s16 q2, q3, d2[2] @weight 2 mult. for row 1H vmovl.u8 q4, d9 @converting row 2H in source 1 to 16-bit vmovl.u8 q5, d11 @converting row 2H in source 2 to 16-bit vmul.s16 q12, q12, d2[0] @weight 1 mult. for row 2L vmla.s16 q12, q13, d2[2] @weight 2 mult. for row 2L vmovl.u8 q14, d12 @converting row 3L in source 1 to 16-bit vmovl.u8 q15, d14 @converting row 3L in source 2 to 16-bit vmul.s16 q4, q4, d2[0] @weight 1 mult. for row 2H vmla.s16 q4, q5, d2[2] @weight 2 mult. for row 2H vmovl.u8 q6, d13 @converting row 3H in source 1 to 16-bit vmovl.u8 q7, d15 @converting row 3H in source 2 to 16-bit vmul.s16 q14, q14, d2[0] @weight 1 mult. for row 3L vmla.s16 q14, q15, d2[2] @weight 2 mult. for row 3L vmovl.u8 q11, d16 @converting row 4L in source 1 to 16-bit vmovl.u8 q3, d18 @converting row 4L in source 2 to 16-bit vmul.s16 q6, q6, d2[0] @weight 1 mult. for row 3H vmla.s16 q6, q7, d2[2] @weight 2 mult. for row 3H vmovl.u8 q8, d17 @converting row 4H in source 1 to 16-bit vmovl.u8 q9, d19 @converting row 4H in source 2 to 16-bit vmul.s16 q11, q11, d2[0] @weight 1 mult. for row 4L vmla.s16 q11, q3, d2[2] @weight 2 mult. for row 4L vrshl.s16 q10, q10, q0 @rounds off the weighted samples from row 1L vmul.s16 q8, q8, d2[0] @weight 1 mult. for row 4H vmla.s16 q8, q9, d2[2] @weight 2 mult. for row 4H vrshl.s16 q2, q2, q0 @rounds off the weighted samples from row 1H vrshl.s16 q12, q12, q0 @rounds off the weighted samples from row 2L vaddw.s8 q10, q10, d3 @adding offset for row 1L vrshl.s16 q4, q4, q0 @rounds off the weighted samples from row 2H vaddw.s8 q2, q2, d3 @adding offset for row 1H vrshl.s16 q14, q14, q0 @rounds off the weighted samples from row 3L vaddw.s8 q12, q12, d3 @adding offset for row 2L vrshl.s16 q6, q6, q0 @rounds off the weighted samples from row 3H vaddw.s8 q4, q4, d3 @adding offset for row 2H vrshl.s16 q11, q11, q0 @rounds off the weighted samples from row 4L vaddw.s8 q14, q14, d3 @adding offset for row 3L vrshl.s16 q8, q8, q0 @rounds off the weighted samples from row 4H vaddw.s8 q6, q6, d3 @adding offset for row 3H vqmovun.s16 d26, q10 @saturating row 1L to unsigned 8-bit vaddw.s8 q11, q11, d3 @adding offset for row 4L vqmovun.s16 d27, q2 @saturating row 1H to unsigned 8-bit vaddw.s8 q8, q8, d3 @adding offset for row 4H vqmovun.s16 d10, q12 @saturating row 2L to unsigned 8-bit vqmovun.s16 d11, q4 @saturating row 2H to unsigned 8-bit vqmovun.s16 d30, q14 @saturating row 3L to unsigned 8-bit vqmovun.s16 d31, q6 @saturating row 3H to unsigned 8-bit vst1.8 {q13}, [r2], r5 @store row 1 in destination vqmovun.s16 d14, q11 @saturating row 4L to unsigned 8-bit vqmovun.s16 d15, q8 @saturating row 4H to unsigned 8-bit vst1.8 {q5}, [r2], r5 @store row 2 in destination subs r11, r11, #4 @decrement ht by 4 vst1.8 {q15}, [r2], r5 @store row 3 in destination vst1.8 {q7}, [r2], r5 @store row 4 in destination bgt loop_16 @if greater than 0 repeat the loop again end_loops: vpop {d8-d15} ldmfd sp!, {r4-r12, r15} @Reload the registers from sp @******************************************************************************* @* @function @* ih264_weighted_bi_pred_chroma_a9q() @* @* @brief @* This routine performs the default weighted prediction as described in sec @* 8.4.2.3.2 titled "Weighted sample prediction process" for chroma. @* @* @par Description: @* This function gets two ht x wd blocks, calculates the weighted samples, @* rounds off, adds offset and stores it in the destination block for U and V. @* @* @param[in] pu1_src1 @* UWORD8 Pointer to the buffer containing the input block 1. @* @* @param[in] pu1_src2 @* UWORD8 Pointer to the buffer containing the input block 2. @* @* @param[out] pu1_dst @* UWORD8 pointer to the destination where the output block is stored. @* @* @param[in] src_strd1 @* Stride of the input buffer 1 @* @* @param[in] src_strd2 @* Stride of the input buffer 2 @* @* @param[in] dst_strd @* Stride of the destination buffer @* @* @param[in] log_wd @* number of bits to be rounded off @* @* @param[in] wt1 @* weights for the weighted prediction in U and V @* @* @param[in] wt2 @* weights for the weighted prediction in U and V @* @* @param[in] ofst1 @* offset 1 used after rounding off for U an dV @* @* @param[in] ofst2 @* offset 2 used after rounding off for U and V @* @* @param[in] ht @* integer height of the array @* @* @param[in] wd @* integer width of the array @* @* @returns @* None @* @* @remarks @* (ht,wd) can be (2,2), (2,4), (4,2), (4,4), (4,8), (8,4) or (8,8). @* @******************************************************************************* @* @void ih264_weighted_bi_pred_chroma_a9q(UWORD8 *pu1_src1, @ UWORD8 *pu1_src2, @ UWORD8 *pu1_dst, @ WORD32 src_strd1, @ WORD32 src_strd2, @ WORD32 dst_strd, @ WORD32 log_wd, @ WORD32 wt1, @ WORD32 wt2, @ WORD32 ofst1, @ WORD32 ofst2, @ WORD32 ht, @ WORD32 wd) @ @**************Variables Vs Registers***************************************** @ r0 => pu1_src1 @ r1 => pu1_src2 @ r2 => pu1_dst @ r3 => src_strd1 @ [sp] => src_strd2 (r4) @ [sp+4] => dst_strd (r5) @ [sp+8] => log_wd (r6) @ [sp+12] => wt1 (r7) @ [sp+16] => wt2 (r8) @ [sp+20] => ofst1 (r9) @ [sp+24] => ofst2 (r10) @ [sp+28] => ht (r11) @ [sp+32] => wd (r12) @ .global ih264_weighted_bi_pred_chroma_a9q ih264_weighted_bi_pred_chroma_a9q: stmfd sp!, {r4-r12, r14} @stack stores the values of the arguments ldr r6, [sp, #48] @Load log_wd in r6 ldr r7, [sp, #52] @Load wt1 in r7 ldr r8, [sp, #56] @Load wt2 in r8 add r6, r6, #1 @r6 = log_wd + 1 ldr r9, [sp, #60] @Load ofst1 in r9 ldr r10, [sp, #64] @Load ofst2 in r10 neg r12, r6 @r12 = -(log_wd + 1) ldr r4, [sp, #40] @Load src_strd2 in r4 ldr r5, [sp, #44] @Load dst_strd in r5 vdup.16 q0, r12 @Q0 = -(log_wd + 1) (16-bit) ldr r11, [sp, #68] @Load ht in r11 vdup.32 q1, r7 @Q1 = (wt1_u, wt1_v) (32-bit) ldr r12, [sp, #72] @Load wd in r12 vdup.32 q2, r8 @Q2 = (wt2_u, wt2_v) (32-bit) asr r7, r9, #8 @r7 = ofst1_v asr r8, r10, #8 @r8 = ofst2_v vpush {d8-d15} sxtb r9, r9 @sign-extend 8-bit ofst1_u to 32-bit sxtb r10, r10 @sign-extend 8-bit ofst2_u to 32-bit sxtb r7, r7 @sign-extend 8-bit ofst1_v to 32-bit sxtb r8, r8 @sign-extend 8-bit ofst2_v to 32-bit add r9, r9, #1 @r9 = ofst1_u + 1 add r7, r7, #1 @r7 = ofst1_v + 1 add r9, r9, r10 @r9 = ofst1_u + ofst2_u + 1 add r7, r7, r8 @r7 = ofst1_v + ofst2_v + 1 asr r9, r9, #1 @r9 = ofst_u = (ofst1_u + ofst2_u + 1) >> 1 asr r7, r7, #1 @r7 = ofst_v = (ofst1_v + ofst2_v + 1) >> 1 cmp r12, #8 @check if wd is 8 pkhbt r9, r9, r7, lsl #16 @r9 = {ofst_u(16-bit), ofst_v(16-bit)} vdup.32 q3, r9 @Q3 = {ofst_u(16-bit), ofst_v(16-bit)} beq loop_8_uv @branch if wd is 8 cmp r12, #4 @check if wd is 4 beq loop_4_uv @branch if wd is 4 loop_2_uv: @each iteration processes two rows vld1.32 d8[0], [r0], r3 @load row 1 in source 1 vld1.32 d8[1], [r0], r3 @load row 2 in source 1 vld1.32 d10[0], [r1], r4 @load row 1 in source 2 vld1.32 d10[1], [r1], r4 @load row 2 in source 2 vmovl.u8 q4, d8 @converting rows 1,2 in source 1 to 16-bit vmovl.u8 q5, d10 @converting rows 1,2 in source 2 to 16-bit vmul.s16 q4, q4, q1 @weight 1 mult. for rows 1,2 vmla.s16 q4, q5, q2 @weight 2 mult. for rows 1,2 vrshl.s16 q4, q4, q0 @rounds off the weighted samples from rows 1,2 vadd.s16 q4, q4, q3 @adding offset for rows 1,2 vqmovun.s16 d8, q4 @saturating rows 1,2 to unsigned 8-bit vst1.32 d8[0], [r2], r5 @store row 1 in destination vst1.32 d8[1], [r2], r5 @store row 2 in destination subs r11, r11, #2 @decrement ht by 2 bgt loop_2_uv @if greater than 0 repeat the loop again b end_loops_uv loop_4_uv: @each iteration processes two rows vld1.8 d8, [r0], r3 @load row 1 in source 1 vld1.8 d10, [r1], r4 @load row 1 in source 2 vmovl.u8 q4, d8 @converting row 1 in source 1 to 16-bit vld1.8 d12, [r0], r3 @load row 2 in source 1 vmovl.u8 q5, d10 @converting row 1 in source 2 to 16-bit vld1.8 d14, [r1], r4 @load row 2 in source 2 vmovl.u8 q6, d12 @converting row 2 in source 1 to 16-bit vmul.s16 q4, q4, q1 @weight 1 mult. for row 1 vmla.s16 q4, q5, q2 @weight 2 mult. for row 1 vmovl.u8 q7, d14 @converting row 2 in source 2 to 16-bit vmul.s16 q6, q6, q1 @weight 1 mult. for row 2 vmla.s16 q6, q7, q2 @weight 2 mult. for row 2 subs r11, r11, #2 @decrement ht by 2 vrshl.s16 q4, q4, q0 @rounds off the weighted samples from row 1 vrshl.s16 q6, q6, q0 @rounds off the weighted samples from row 2 vadd.s16 q4, q4, q3 @adding offset for row 1 vadd.s16 q6, q6, q3 @adding offset for row 2 vqmovun.s16 d8, q4 @saturating row 1 to unsigned 8-bit vqmovun.s16 d12, q6 @saturating row 2 to unsigned 8-bit vst1.8 d8, [r2], r5 @store row 1 in destination vst1.8 d12, [r2], r5 @store row 2 in destination bgt loop_4_uv @if greater than 0 repeat the loop again b end_loops_uv loop_8_uv: @each iteration processes two rows vld1.8 {q4}, [r0], r3 @load row 1 in source 1 vld1.8 {q5}, [r1], r4 @load row 1 in source 2 vld1.8 {q6}, [r0], r3 @load row 2 in source 1 vld1.8 {q7}, [r1], r4 @load row 2 in source 2 vmovl.u8 q12, d8 @converting row 1L in source 1 to 16-bit vld1.8 {q8}, [r0], r3 @load row 3 in source 1 vld1.8 {q9}, [r1], r4 @load row 3 in source 2 vmovl.u8 q13, d10 @converting row 1L in source 2 to 16-bit vld1.8 {q10}, [r0], r3 @load row 4 in source 1 vld1.8 {q11}, [r1], r4 @load row 4 in source 2 vmovl.u8 q4, d9 @converting row 1H in source 1 to 16-bit vmovl.u8 q5, d11 @converting row 1H in source 2 to 16-bit vmul.s16 q12, q12, q1 @weight 1 mult. for row 1L vmla.s16 q12, q13, q2 @weight 2 mult. for row 1L vmovl.u8 q14, d12 @converting row 2L in source 1 to 16-bit vmovl.u8 q15, d14 @converting row 2L in source 2 to 16-bit vmul.s16 q4, q4, q1 @weight 1 mult. for row 1H vmla.s16 q4, q5, q2 @weight 2 mult. for row 1H vmovl.u8 q6, d13 @converting row 2H in source 1 to 16-bit vmovl.u8 q7, d15 @converting row 2H in source 2 to 16-bit vmul.s16 q14, q14, q1 @weight 1 mult. for row 2L vmla.s16 q14, q15, q2 @weight 2 mult. for row 2L vmovl.u8 q13, d16 @converting row 3L in source 1 to 16-bit vmovl.u8 q5, d18 @converting row 3L in source 2 to 16-bit vmul.s16 q6, q6, q1 @weight 1 mult. for row 2H vmla.s16 q6, q7, q2 @weight 2 mult. for row 2H vmovl.u8 q8, d17 @converting row 3H in source 1 to 16-bit vmovl.u8 q9, d19 @converting row 3H in source 2 to 16-bit vmul.s16 q13, q13, q1 @weight 1 mult. for row 3L vmla.s16 q13, q5, q2 @weight 2 mult. for row 3L vmovl.u8 q15, d20 @converting row 4L in source 1 to 16-bit vmovl.u8 q7, d22 @converting row 4L in source 2 to 16-bit vmul.s16 q8, q8, q1 @weight 1 mult. for row 3H vmla.s16 q8, q9, q2 @weight 2 mult. for row 3H vmovl.u8 q10, d21 @converting row 4H in source 1 to 16-bit vmovl.u8 q11, d23 @converting row 4H in source 2 to 16-bit vmul.s16 q15, q15, q1 @weight 1 mult. for row 4L vmla.s16 q15, q7, q2 @weight 2 mult. for row 4L vrshl.s16 q12, q12, q0 @rounds off the weighted samples from row 1L vmul.s16 q10, q10, q1 @weight 1 mult. for row 4H vmla.s16 q10, q11, q2 @weight 2 mult. for row 4H vrshl.s16 q4, q4, q0 @rounds off the weighted samples from row 1H vrshl.s16 q14, q14, q0 @rounds off the weighted samples from row 2L vadd.s16 q12, q12, q3 @adding offset for row 1L vrshl.s16 q6, q6, q0 @rounds off the weighted samples from row 2H vadd.s16 q4, q4, q3 @adding offset for row 1H vrshl.s16 q13, q13, q0 @rounds off the weighted samples from row 3L vadd.s16 q14, q14, q3 @adding offset for row 2L vrshl.s16 q8, q8, q0 @rounds off the weighted samples from row 3H vadd.s16 q6, q6, q3 @adding offset for row 2H vrshl.s16 q15, q15, q0 @rounds off the weighted samples from row 4L vadd.s16 q13, q13, q3 @adding offset for row 3L vrshl.s16 q10, q10, q0 @rounds off the weighted samples from row 4H vadd.s16 q8, q8, q3 @adding offset for row 3H vqmovun.s16 d10, q12 @saturating row 1L to unsigned 8-bit vadd.s16 q15, q15, q3 @adding offset for row 4L vqmovun.s16 d11, q4 @saturating row 1H to unsigned 8-bit vadd.s16 q10, q10, q3 @adding offset for row 4H vqmovun.s16 d18, q14 @saturating row 2L to unsigned 8-bit vqmovun.s16 d19, q6 @saturating row 2H to unsigned 8-bit vqmovun.s16 d14, q13 @saturating row 3L to unsigned 8-bit vqmovun.s16 d15, q8 @saturating row 3H to unsigned 8-bit vst1.8 {q5}, [r2], r5 @store row 1 in destination vqmovun.s16 d22, q15 @saturating row 4L to unsigned 8-bit vqmovun.s16 d23, q10 @saturating row 4H to unsigned 8-bit vst1.8 {q9}, [r2], r5 @store row 2 in destination subs r11, r11, #4 @decrement ht by 4 vst1.8 {q7}, [r2], r5 @store row 3 in destination vst1.8 {q11}, [r2], r5 @store row 4 in destination bgt loop_8_uv @if greater than 0 repeat the loop again end_loops_uv: vpop {d8-d15} ldmfd sp!, {r4-r12, r15} @Reload the registers from sp