yading@10: /* yading@10: * Copyright (c) 2011 Janne Grunau yading@10: * Copyright (c) 2011 Mans Rullgard yading@10: * yading@10: * This file is part of Libav. yading@10: * yading@10: * Libav is free software; you can redistribute it and/or yading@10: * modify it under the terms of the GNU Lesser General Public yading@10: * License as published by the Free Software Foundation; either yading@10: * version 2.1 of the License, or (at your option) any later version. yading@10: * yading@10: * Libav is distributed in the hope that it will be useful, yading@10: * but WITHOUT ANY WARRANTY; without even the implied warranty of yading@10: * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU yading@10: * Lesser General Public License for more details. yading@10: * yading@10: * You should have received a copy of the GNU Lesser General Public yading@10: * License along with Libav; if not, write to the Free Software yading@10: * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA yading@10: */ yading@10: yading@10: #include "libavutil/arm/asm.S" yading@10: #include "neon.S" yading@10: yading@10: .macro qpel_lowpass r0, r1, rc1, rc2, shift yading@10: vext.8 d25, \r0, \r1, #1 @ src[-1] yading@10: vext.8 d26, \r0, \r1, #4 @ src[ 2] yading@10: vext.8 d24, \r0, \r1, #5 @ src[ 3] yading@10: vaddl.u8 q9, d25, d26 yading@10: vaddl.u8 q8, \r0, d24 yading@10: vext.8 d27, \r0, \r1, #2 @ src[ 0] yading@10: vshl.s16 q12, q9, #2 yading@10: vsub.s16 q8, q8, q9 yading@10: vext.8 d28, \r0, \r1, #3 @ src[ 1] yading@10: vsub.s16 q8, q8, q12 yading@10: vmlal.u8 q8, d27, \rc1 yading@10: vmlal.u8 q8, d28, \rc2 yading@10: vqrshrun.s16 \r0, q8, #\shift yading@10: .endm yading@10: yading@10: .macro qpel_lowpass_x2 r0, r1, r2, r3, rc1, rc2, shift yading@10: vext.8 d25, \r0, \r1, #1 @ src[-1] yading@10: vext.8 d26, \r0, \r1, #4 @ src[ 2] yading@10: vext.8 d24, \r0, \r1, #5 @ src[ 3] yading@10: vaddl.u8 q9, d25, d26 yading@10: vaddl.u8 q8, \r0, d24 yading@10: vext.8 d29, \r0, \r1, #2 @ src[ 0] yading@10: vext.8 d28, \r0, \r1, #3 @ src[ 1] yading@10: vshl.s16 q10, q9, #2 yading@10: vext.8 \r1, \r2, \r3, #1 @ src[-1] yading@10: vsub.s16 q8, q8, q9 yading@10: vext.8 d22, \r2, \r3, #4 @ src[ 2] yading@10: vext.8 \r0, \r2, \r3, #5 @ src[ 3] yading@10: vaddl.u8 q13, \r1, d22 yading@10: vaddl.u8 q12, \r2, \r0 yading@10: vsub.s16 q8, q8, q10 yading@10: vshl.s16 q9, q13, #2 yading@10: vsub.s16 q12, q12, q13 yading@10: vmlal.u8 q8, d29, \rc1 yading@10: vmlal.u8 q8, d28, \rc2 yading@10: vsub.s16 q12, q12, q9 yading@10: vext.8 d26, \r2, \r3, #2 @ src[ 0] yading@10: vext.8 d27, \r2, \r3, #3 @ src[ 1] yading@10: vmlal.u8 q12, d26, \rc1 yading@10: vmlal.u8 q12, d27, \rc2 yading@10: vqrshrun.s16 \r0, q8, #\shift yading@10: vqrshrun.s16 \r2, q12, #\shift yading@10: .endm yading@10: yading@10: .macro rv40_qpel8_h shift yading@10: function put_rv40_qpel8_h_lp_packed_s\shift\()_neon yading@10: 1: yading@10: vld1.8 {q2}, [r1], r2 yading@10: vld1.8 {q3}, [r1], r2 yading@10: qpel_lowpass_x2 d4, d5, d6, d7, d0, d1, \shift yading@10: vst1.8 {d4}, [r12,:64]! yading@10: vst1.8 {d6}, [r12,:64]! yading@10: subs r3, r3, #2 yading@10: bgt 1b yading@10: vld1.8 {q2}, [r1] yading@10: qpel_lowpass d4, d5, d0, d1, \shift yading@10: vst1.8 {d4}, [r12,:64]! yading@10: bx lr yading@10: endfunc yading@10: .endm yading@10: yading@10: .macro rv40_qpel8_v shift, type yading@10: function \type\()_rv40_qpel8_v_lp_packed_s\shift\()_neon yading@10: vld1.64 {d2}, [r1,:64]! yading@10: vld1.64 {d3}, [r1,:64]! yading@10: vld1.64 {d4}, [r1,:64]! yading@10: vld1.64 {d5}, [r1,:64]! yading@10: vld1.64 {d6}, [r1,:64]! yading@10: vld1.64 {d7}, [r1,:64]! yading@10: vld1.64 {d8}, [r1,:64]! yading@10: vld1.64 {d9}, [r1,:64]! yading@10: vld1.64 {d10}, [r1,:64]! yading@10: vld1.64 {d11}, [r1,:64]! yading@10: vld1.64 {d12}, [r1,:64]! yading@10: vld1.64 {d13}, [r1,:64]! yading@10: vld1.64 {d14}, [r1,:64]! yading@10: transpose_8x8 d2, d3, d4, d5, d6, d7, d8, d9 yading@10: transpose_8x8 d10, d11, d12, d13, d14, d15, d30, d31 yading@10: qpel_lowpass_x2 d2, d10, d3, d11, d0, d1, \shift yading@10: qpel_lowpass_x2 d4, d12, d5, d13, d0, d1, \shift yading@10: qpel_lowpass_x2 d6, d14, d7, d15, d0, d1, \shift yading@10: qpel_lowpass_x2 d8, d30, d9, d31, d0, d1, \shift yading@10: transpose_8x8 d2, d3, d4, d5, d6, d7, d8, d9 yading@10: .ifc \type,avg yading@10: vld1.64 d12, [r0,:64], r2 yading@10: vld1.64 d13, [r0,:64], r2 yading@10: vld1.64 d14, [r0,:64], r2 yading@10: vld1.64 d15, [r0,:64], r2 yading@10: vld1.64 d16, [r0,:64], r2 yading@10: vld1.64 d17, [r0,:64], r2 yading@10: vld1.64 d18, [r0,:64], r2 yading@10: vld1.64 d19, [r0,:64], r2 yading@10: sub r0, r0, r2, lsl #3 yading@10: vrhadd.u8 q1, q1, q6 yading@10: vrhadd.u8 q2, q2, q7 yading@10: vrhadd.u8 q3, q3, q8 yading@10: vrhadd.u8 q4, q4, q9 yading@10: .endif yading@10: vst1.64 d2, [r0,:64], r2 yading@10: vst1.64 d3, [r0,:64], r2 yading@10: vst1.64 d4, [r0,:64], r2 yading@10: vst1.64 d5, [r0,:64], r2 yading@10: vst1.64 d6, [r0,:64], r2 yading@10: vst1.64 d7, [r0,:64], r2 yading@10: vst1.64 d8, [r0,:64], r2 yading@10: vst1.64 d9, [r0,:64], r2 yading@10: bx lr yading@10: endfunc yading@10: .endm yading@10: yading@10: rv40_qpel8_h 5 yading@10: rv40_qpel8_h 6 yading@10: yading@10: .macro rv40_qpel type yading@10: function \type\()_rv40_qpel8_h_lowpass_neon yading@10: .ifc \type,avg yading@10: mov r12, r0 yading@10: .endif yading@10: 1: yading@10: vld1.8 {q2}, [r1], r2 yading@10: vld1.8 {q3}, [r1], r2 yading@10: qpel_lowpass_x2 d4, d5, d6, d7, d0, d1, 6 yading@10: .ifc \type,avg yading@10: vld1.8 {d3}, [r12,:64], r2 yading@10: vld1.8 {d16}, [r12,:64], r2 yading@10: vrhadd.u8 d4, d4, d3 yading@10: vrhadd.u8 d6, d6, d16 yading@10: .endif yading@10: vst1.8 {d4}, [r0,:64], r2 yading@10: vst1.8 {d6}, [r0,:64], r2 yading@10: subs r3, r3, #2 yading@10: bgt 1b yading@10: bx lr yading@10: endfunc yading@10: yading@10: function \type\()_rv40_qpel8_v_lowpass_neon yading@10: vld1.64 {d2}, [r1], r2 yading@10: vld1.64 {d3}, [r1], r2 yading@10: vld1.64 {d4}, [r1], r2 yading@10: vld1.64 {d5}, [r1], r2 yading@10: vld1.64 {d6}, [r1], r2 yading@10: vld1.64 {d7}, [r1], r2 yading@10: vld1.64 {d8}, [r1], r2 yading@10: vld1.64 {d9}, [r1], r2 yading@10: vld1.64 {d10}, [r1], r2 yading@10: vld1.64 {d11}, [r1], r2 yading@10: vld1.64 {d12}, [r1], r2 yading@10: vld1.64 {d13}, [r1], r2 yading@10: vld1.64 {d14}, [r1] yading@10: transpose_8x8 d2, d3, d4, d5, d6, d7, d8, d9 yading@10: transpose_8x8 d10, d11, d12, d13, d14, d15, d30, d31 yading@10: qpel_lowpass_x2 d2, d10, d3, d11, d0, d1, 6 yading@10: qpel_lowpass_x2 d4, d12, d5, d13, d0, d1, 6 yading@10: qpel_lowpass_x2 d6, d14, d7, d15, d0, d1, 6 yading@10: qpel_lowpass_x2 d8, d30, d9, d31, d0, d1, 6 yading@10: transpose_8x8 d2, d3, d4, d5, d6, d7, d8, d9 yading@10: .ifc \type,avg yading@10: vld1.64 d12, [r0,:64], r2 yading@10: vld1.64 d13, [r0,:64], r2 yading@10: vld1.64 d14, [r0,:64], r2 yading@10: vld1.64 d15, [r0,:64], r2 yading@10: vld1.64 d16, [r0,:64], r2 yading@10: vld1.64 d17, [r0,:64], r2 yading@10: vld1.64 d18, [r0,:64], r2 yading@10: vld1.64 d19, [r0,:64], r2 yading@10: sub r0, r0, r2, lsl #3 yading@10: vrhadd.u8 q1, q1, q6 yading@10: vrhadd.u8 q2, q2, q7 yading@10: vrhadd.u8 q3, q3, q8 yading@10: vrhadd.u8 q4, q4, q9 yading@10: .endif yading@10: vst1.64 d2, [r0,:64], r2 yading@10: vst1.64 d3, [r0,:64], r2 yading@10: vst1.64 d4, [r0,:64], r2 yading@10: vst1.64 d5, [r0,:64], r2 yading@10: vst1.64 d6, [r0,:64], r2 yading@10: vst1.64 d7, [r0,:64], r2 yading@10: vst1.64 d8, [r0,:64], r2 yading@10: vst1.64 d9, [r0,:64], r2 yading@10: bx lr yading@10: endfunc yading@10: yading@10: rv40_qpel8_v 5, \type yading@10: rv40_qpel8_v 6, \type yading@10: yading@10: function ff_\type\()_rv40_qpel8_mc10_neon, export=1 yading@10: sub r1, r1, #2 yading@10: mov r3, #8 yading@10: vmov.i8 d0, #52 yading@10: vmov.i8 d1, #20 yading@10: b \type\()_rv40_qpel8_h_lowpass_neon yading@10: endfunc yading@10: yading@10: function ff_\type\()_rv40_qpel8_mc30_neon, export=1 yading@10: sub r1, r1, #2 yading@10: mov r3, #8 yading@10: vmov.i8 d0, #20 yading@10: vmov.i8 d1, #52 yading@10: b \type\()_rv40_qpel8_h_lowpass_neon yading@10: endfunc yading@10: yading@10: function ff_\type\()_rv40_qpel8_mc01_neon, export=1 yading@10: push {r4, lr} yading@10: vpush {d8-d15} yading@10: sub r1, r1, r2, lsl #1 yading@10: vmov.i8 d0, #52 yading@10: vmov.i8 d1, #20 yading@10: bl \type\()_rv40_qpel8_v_lowpass_neon yading@10: vpop {d8-d15} yading@10: pop {r4, pc} yading@10: endfunc yading@10: yading@10: function ff_\type\()_rv40_qpel8_mc11_neon, export=1 yading@10: push {r4, lr} yading@10: vpush {d8-d15} yading@10: sub sp, sp, #14*8 yading@10: add r12, sp, #7 yading@10: bic r12, r12, #7 yading@10: sub r1, r1, r2, lsl #1 yading@10: sub r1, r1, #2 yading@10: mov r3, #12 yading@10: vmov.i8 d0, #52 yading@10: vmov.i8 d1, #20 yading@10: bl put_rv40_qpel8_h_lp_packed_s6_neon yading@10: add r1, sp, #7 yading@10: bic r1, r1, #7 yading@10: bl \type\()_rv40_qpel8_v_lp_packed_s6_neon yading@10: add sp, sp, #14*8 yading@10: vpop {d8-d15} yading@10: pop {r4, pc} yading@10: endfunc yading@10: yading@10: function ff_\type\()_rv40_qpel8_mc21_neon, export=1 yading@10: push {r4, lr} yading@10: vpush {d8-d15} yading@10: sub sp, sp, #14*8 yading@10: add r12, sp, #7 yading@10: bic r12, r12, #7 yading@10: sub r1, r1, r2, lsl #1 yading@10: sub r1, r1, #2 yading@10: mov r3, #12 yading@10: vmov.i8 d0, #20 yading@10: vmov.i8 d1, #20 yading@10: bl put_rv40_qpel8_h_lp_packed_s5_neon yading@10: add r1, sp, #7 yading@10: bic r1, r1, #7 yading@10: vmov.i8 d0, #52 yading@10: bl \type\()_rv40_qpel8_v_lp_packed_s6_neon yading@10: add sp, sp, #14*8 yading@10: vpop {d8-d15} yading@10: pop {r4, pc} yading@10: endfunc yading@10: yading@10: function ff_\type\()_rv40_qpel8_mc31_neon, export=1 yading@10: push {r4, lr} yading@10: vpush {d8-d15} yading@10: sub sp, sp, #14*8 yading@10: add r12, sp, #7 yading@10: bic r12, r12, #7 yading@10: sub r1, r1, r2, lsl #1 yading@10: sub r1, r1, #2 yading@10: mov r3, #12 yading@10: vmov.i8 d0, #20 yading@10: vmov.i8 d1, #52 yading@10: bl put_rv40_qpel8_h_lp_packed_s6_neon yading@10: add r1, sp, #7 yading@10: bic r1, r1, #7 yading@10: vswp d0, d1 yading@10: bl \type\()_rv40_qpel8_v_lp_packed_s6_neon yading@10: add sp, sp, #14*8 yading@10: vpop {d8-d15} yading@10: pop {r4, pc} yading@10: endfunc yading@10: yading@10: function ff_\type\()_rv40_qpel8_mc12_neon, export=1 yading@10: push {r4, lr} yading@10: vpush {d8-d15} yading@10: sub sp, sp, #14*8 yading@10: add r12, sp, #7 yading@10: bic r12, r12, #7 yading@10: sub r1, r1, r2, lsl #1 yading@10: sub r1, r1, #2 yading@10: mov r3, #12 yading@10: vmov.i8 d0, #52 yading@10: vmov.i8 d1, #20 yading@10: bl put_rv40_qpel8_h_lp_packed_s6_neon yading@10: add r1, sp, #7 yading@10: bic r1, r1, #7 yading@10: vmov.i8 d0, #20 yading@10: bl \type\()_rv40_qpel8_v_lp_packed_s5_neon yading@10: add sp, sp, #14*8 yading@10: vpop {d8-d15} yading@10: pop {r4, pc} yading@10: endfunc yading@10: yading@10: function ff_\type\()_rv40_qpel8_mc22_neon, export=1 yading@10: push {r4, lr} yading@10: vpush {d8-d15} yading@10: sub sp, sp, #14*8 yading@10: add r12, sp, #7 yading@10: bic r12, r12, #7 yading@10: sub r1, r1, r2, lsl #1 yading@10: sub r1, r1, #2 yading@10: mov r3, #12 yading@10: vmov.i8 d0, #20 yading@10: vmov.i8 d1, #20 yading@10: bl put_rv40_qpel8_h_lp_packed_s5_neon yading@10: add r1, sp, #7 yading@10: bic r1, r1, #7 yading@10: bl \type\()_rv40_qpel8_v_lp_packed_s5_neon yading@10: add sp, sp, #14*8 yading@10: vpop {d8-d15} yading@10: pop {r4, pc} yading@10: endfunc yading@10: yading@10: function ff_\type\()_rv40_qpel8_mc32_neon, export=1 yading@10: push {r4, lr} yading@10: vpush {d8-d15} yading@10: sub sp, sp, #14*8 yading@10: add r12, sp, #7 yading@10: bic r12, r12, #7 yading@10: sub r1, r1, r2, lsl #1 yading@10: sub r1, r1, #2 yading@10: mov r3, #12 yading@10: vmov.i8 d0, #20 yading@10: vmov.i8 d1, #52 yading@10: bl put_rv40_qpel8_h_lp_packed_s6_neon yading@10: add r1, sp, #7 yading@10: bic r1, r1, #7 yading@10: vmov.i8 d1, #20 yading@10: bl \type\()_rv40_qpel8_v_lp_packed_s5_neon yading@10: add sp, sp, #14*8 yading@10: vpop {d8-d15} yading@10: pop {r4, pc} yading@10: endfunc yading@10: yading@10: function ff_\type\()_rv40_qpel8_mc03_neon, export=1 yading@10: push {r4, lr} yading@10: vpush {d8-d15} yading@10: sub r1, r1, r2, lsl #1 yading@10: vmov.i8 d0, #20 yading@10: vmov.i8 d1, #52 yading@10: bl \type\()_rv40_qpel8_v_lowpass_neon yading@10: vpop {d8-d15} yading@10: pop {r4, pc} yading@10: endfunc yading@10: yading@10: function ff_\type\()_rv40_qpel8_mc33_neon, export=1 yading@10: mov r3, #8 yading@10: b X(ff_\type\()_pixels8_xy2_neon) yading@10: endfunc yading@10: yading@10: function ff_\type\()_rv40_qpel8_mc13_neon, export=1 yading@10: push {r4, lr} yading@10: vpush {d8-d15} yading@10: sub sp, sp, #14*8 yading@10: add r12, sp, #7 yading@10: bic r12, r12, #7 yading@10: sub r1, r1, r2, lsl #1 yading@10: sub r1, r1, #2 yading@10: mov r3, #12 yading@10: vmov.i8 d0, #52 yading@10: vmov.i8 d1, #20 yading@10: bl put_rv40_qpel8_h_lp_packed_s6_neon yading@10: add r1, sp, #7 yading@10: bic r1, r1, #7 yading@10: vswp d0, d1 yading@10: bl \type\()_rv40_qpel8_v_lp_packed_s6_neon yading@10: add sp, sp, #14*8 yading@10: vpop {d8-d15} yading@10: pop {r4, pc} yading@10: endfunc yading@10: yading@10: function ff_\type\()_rv40_qpel8_mc23_neon, export=1 yading@10: push {r4, lr} yading@10: vpush {d8-d15} yading@10: sub sp, sp, #14*8 yading@10: add r12, sp, #7 yading@10: bic r12, r12, #7 yading@10: sub r1, r1, r2, lsl #1 yading@10: sub r1, r1, #2 yading@10: mov r3, #12 yading@10: vmov.i8 d0, #20 yading@10: vmov.i8 d1, #20 yading@10: bl put_rv40_qpel8_h_lp_packed_s5_neon yading@10: add r1, sp, #7 yading@10: bic r1, r1, #7 yading@10: vmov.i8 d1, #52 yading@10: bl \type\()_rv40_qpel8_v_lp_packed_s6_neon yading@10: add sp, sp, #14*8 yading@10: vpop {d8-d15} yading@10: pop {r4, pc} yading@10: endfunc yading@10: yading@10: function ff_\type\()_rv40_qpel16_mc10_neon, export=1 yading@10: vmov.i8 d0, #52 yading@10: vmov.i8 d1, #20 yading@10: .L\type\()_rv40_qpel16_h: yading@10: push {r1, lr} yading@10: sub r1, r1, #2 yading@10: mov r3, #16 yading@10: bl \type\()_rv40_qpel8_h_lowpass_neon yading@10: pop {r1, lr} yading@10: sub r0, r0, r2, lsl #4 yading@10: add r0, r0, #8 yading@10: add r1, r1, #6 yading@10: mov r3, #16 yading@10: b \type\()_rv40_qpel8_h_lowpass_neon yading@10: endfunc yading@10: yading@10: function ff_\type\()_rv40_qpel16_mc30_neon, export=1 yading@10: vmov.i8 d0, #20 yading@10: vmov.i8 d1, #52 yading@10: b .L\type\()_rv40_qpel16_h yading@10: endfunc yading@10: yading@10: function ff_\type\()_rv40_qpel16_mc01_neon, export=1 yading@10: vmov.i8 d0, #52 yading@10: vmov.i8 d1, #20 yading@10: .L\type\()_rv40_qpel16_v: yading@10: sub r1, r1, r2, lsl #1 yading@10: push {r1, lr} yading@10: vpush {d8-d15} yading@10: bl \type\()_rv40_qpel8_v_lowpass_neon yading@10: sub r1, r1, r2, lsl #2 yading@10: bl \type\()_rv40_qpel8_v_lowpass_neon yading@10: ldr r1, [sp, #64] yading@10: sub r0, r0, r2, lsl #4 yading@10: add r0, r0, #8 yading@10: add r1, r1, #8 yading@10: bl \type\()_rv40_qpel8_v_lowpass_neon yading@10: sub r1, r1, r2, lsl #2 yading@10: bl \type\()_rv40_qpel8_v_lowpass_neon yading@10: vpop {d8-d15} yading@10: pop {r1, pc} yading@10: endfunc yading@10: yading@10: function ff_\type\()_rv40_qpel16_mc11_neon, export=1 yading@10: sub r1, r1, r2, lsl #1 yading@10: sub r1, r1, #2 yading@10: push {r1, lr} yading@10: vpush {d8-d15} yading@10: sub sp, sp, #44*8 yading@10: add r12, sp, #7 yading@10: bic r12, r12, #7 yading@10: mov r3, #20 yading@10: vmov.i8 d0, #52 yading@10: vmov.i8 d1, #20 yading@10: bl put_rv40_qpel8_h_lp_packed_s6_neon yading@10: ldr r1, [sp, #416] yading@10: add r1, r1, #8 yading@10: mov r3, #20 yading@10: bl put_rv40_qpel8_h_lp_packed_s6_neon yading@10: .L\type\()_rv40_qpel16_v_s6: yading@10: add r1, sp, #7 yading@10: bic r1, r1, #7 yading@10: bl \type\()_rv40_qpel8_v_lp_packed_s6_neon yading@10: sub r1, r1, #40 yading@10: bl \type\()_rv40_qpel8_v_lp_packed_s6_neon yading@10: sub r0, r0, r2, lsl #4 yading@10: add r0, r0, #8 yading@10: bl \type\()_rv40_qpel8_v_lp_packed_s6_neon yading@10: sub r1, r1, #40 yading@10: bl \type\()_rv40_qpel8_v_lp_packed_s6_neon yading@10: add sp, sp, #44*8 yading@10: vpop {d8-d15} yading@10: pop {r1, pc} yading@10: endfunc yading@10: yading@10: function ff_\type\()_rv40_qpel16_mc21_neon, export=1 yading@10: sub r1, r1, r2, lsl #1 yading@10: sub r1, r1, #2 yading@10: push {r1, lr} yading@10: vpush {d8-d15} yading@10: sub sp, sp, #44*8 yading@10: add r12, sp, #7 yading@10: bic r12, r12, #7 yading@10: mov r3, #20 yading@10: vmov.i8 d0, #20 yading@10: vmov.i8 d1, #20 yading@10: bl put_rv40_qpel8_h_lp_packed_s5_neon yading@10: ldr r1, [sp, #416] yading@10: add r1, r1, #8 yading@10: mov r3, #20 yading@10: bl put_rv40_qpel8_h_lp_packed_s5_neon yading@10: vmov.i8 d0, #52 yading@10: b .L\type\()_rv40_qpel16_v_s6 yading@10: endfunc yading@10: yading@10: function ff_\type\()_rv40_qpel16_mc31_neon, export=1 yading@10: sub r1, r1, r2, lsl #1 yading@10: sub r1, r1, #2 yading@10: push {r1, lr} yading@10: vpush {d8-d15} yading@10: sub sp, sp, #44*8 yading@10: add r12, sp, #7 yading@10: bic r12, r12, #7 yading@10: mov r3, #20 yading@10: vmov.i8 d0, #20 yading@10: vmov.i8 d1, #52 yading@10: bl put_rv40_qpel8_h_lp_packed_s6_neon yading@10: ldr r1, [sp, #416] yading@10: add r1, r1, #8 yading@10: mov r3, #20 yading@10: bl put_rv40_qpel8_h_lp_packed_s6_neon yading@10: vswp d0, d1 yading@10: b .L\type\()_rv40_qpel16_v_s6 yading@10: endfunc yading@10: yading@10: function ff_\type\()_rv40_qpel16_mc12_neon, export=1 yading@10: sub r1, r1, r2, lsl #1 yading@10: sub r1, r1, #2 yading@10: push {r1, lr} yading@10: vpush {d8-d15} yading@10: sub sp, sp, #44*8 yading@10: add r12, sp, #7 yading@10: bic r12, r12, #7 yading@10: mov r3, #20 yading@10: vmov.i8 d0, #52 yading@10: vmov.i8 d1, #20 yading@10: bl put_rv40_qpel8_h_lp_packed_s6_neon yading@10: ldr r1, [sp, #416] yading@10: add r1, r1, #8 yading@10: mov r3, #20 yading@10: bl put_rv40_qpel8_h_lp_packed_s6_neon yading@10: vmov.i8 d0, #20 yading@10: .L\type\()_rv40_qpel16_v_s5: yading@10: add r1, sp, #7 yading@10: bic r1, r1, #7 yading@10: bl \type\()_rv40_qpel8_v_lp_packed_s5_neon yading@10: sub r1, r1, #40 yading@10: bl \type\()_rv40_qpel8_v_lp_packed_s5_neon yading@10: sub r0, r0, r2, lsl #4 yading@10: add r0, r0, #8 yading@10: bl \type\()_rv40_qpel8_v_lp_packed_s5_neon yading@10: sub r1, r1, #40 yading@10: bl \type\()_rv40_qpel8_v_lp_packed_s5_neon yading@10: add sp, sp, #44*8 yading@10: vpop {d8-d15} yading@10: pop {r1, pc} yading@10: endfunc yading@10: yading@10: function ff_\type\()_rv40_qpel16_mc22_neon, export=1 yading@10: sub r1, r1, r2, lsl #1 yading@10: sub r1, r1, #2 yading@10: push {r1, lr} yading@10: vpush {d8-d15} yading@10: sub sp, sp, #44*8 yading@10: add r12, sp, #7 yading@10: bic r12, r12, #7 yading@10: mov r3, #20 yading@10: vmov.i8 d0, #20 yading@10: vmov.i8 d1, #20 yading@10: bl put_rv40_qpel8_h_lp_packed_s5_neon yading@10: ldr r1, [sp, #416] yading@10: add r1, r1, #8 yading@10: mov r3, #20 yading@10: bl put_rv40_qpel8_h_lp_packed_s5_neon yading@10: b .L\type\()_rv40_qpel16_v_s5 yading@10: endfunc yading@10: yading@10: function ff_\type\()_rv40_qpel16_mc32_neon, export=1 yading@10: sub r1, r1, r2, lsl #1 yading@10: sub r1, r1, #2 yading@10: push {r1, lr} yading@10: vpush {d8-d15} yading@10: sub sp, sp, #44*8 yading@10: add r12, sp, #7 yading@10: bic r12, r12, #7 yading@10: mov r3, #20 yading@10: vmov.i8 d0, #20 yading@10: vmov.i8 d1, #52 yading@10: bl put_rv40_qpel8_h_lp_packed_s6_neon yading@10: ldr r1, [sp, #416] yading@10: add r1, r1, #8 yading@10: mov r3, #20 yading@10: bl put_rv40_qpel8_h_lp_packed_s6_neon yading@10: vmov.i8 d1, #20 yading@10: b .L\type\()_rv40_qpel16_v_s5 yading@10: endfunc yading@10: yading@10: function ff_\type\()_rv40_qpel16_mc03_neon, export=1 yading@10: vmov.i8 d0, #20 yading@10: vmov.i8 d1, #52 yading@10: b .L\type\()_rv40_qpel16_v yading@10: endfunc yading@10: yading@10: function ff_\type\()_rv40_qpel16_mc13_neon, export=1 yading@10: sub r1, r1, r2, lsl #1 yading@10: sub r1, r1, #2 yading@10: push {r1, lr} yading@10: vpush {d8-d15} yading@10: sub sp, sp, #44*8 yading@10: add r12, sp, #7 yading@10: bic r12, r12, #7 yading@10: mov r3, #20 yading@10: vmov.i8 d0, #52 yading@10: vmov.i8 d1, #20 yading@10: bl put_rv40_qpel8_h_lp_packed_s6_neon yading@10: ldr r1, [sp, #416] yading@10: add r1, r1, #8 yading@10: mov r3, #20 yading@10: bl put_rv40_qpel8_h_lp_packed_s6_neon yading@10: vswp d0, d1 yading@10: b .L\type\()_rv40_qpel16_v_s6 yading@10: endfunc yading@10: yading@10: function ff_\type\()_rv40_qpel16_mc23_neon, export=1 yading@10: sub r1, r1, r2, lsl #1 yading@10: sub r1, r1, #2 yading@10: push {r1, lr} yading@10: vpush {d8-d15} yading@10: sub sp, sp, #44*8 yading@10: add r12, sp, #7 yading@10: bic r12, r12, #7 yading@10: mov r3, #20 yading@10: vmov.i8 d0, #20 yading@10: vmov.i8 d1, #20 yading@10: bl put_rv40_qpel8_h_lp_packed_s5_neon yading@10: ldr r1, [sp, #416] yading@10: add r1, r1, #8 yading@10: mov r3, #20 yading@10: bl put_rv40_qpel8_h_lp_packed_s5_neon yading@10: vmov.i8 d1, #52 yading@10: b .L\type\()_rv40_qpel16_v_s6 yading@10: endfunc yading@10: yading@10: function ff_\type\()_rv40_qpel16_mc33_neon, export=1 yading@10: mov r3, #16 yading@10: b X(ff_\type\()_pixels16_xy2_neon) yading@10: endfunc yading@10: .endm yading@10: yading@10: rv40_qpel put yading@10: rv40_qpel avg yading@10: yading@10: .macro rv40_weight yading@10: vmovl.u8 q8, d2 yading@10: vmovl.u8 q9, d3 yading@10: vmovl.u8 q10, d4 yading@10: vmovl.u8 q11, d5 yading@10: vmull.u16 q2, d16, d0[2] yading@10: vmull.u16 q3, d17, d0[2] yading@10: vmull.u16 q8, d18, d0[2] yading@10: vmull.u16 q9, d19, d0[2] yading@10: vmull.u16 q12, d20, d0[0] yading@10: vmull.u16 q13, d21, d0[0] yading@10: vmull.u16 q14, d22, d0[0] yading@10: vmull.u16 q15, d23, d0[0] yading@10: vshrn.i32 d4, q2, #9 yading@10: vshrn.i32 d5, q3, #9 yading@10: vshrn.i32 d6, q8, #9 yading@10: vshrn.i32 d7, q9, #9 yading@10: vshrn.i32 d16, q12, #9 yading@10: vshrn.i32 d17, q13, #9 yading@10: vshrn.i32 d18, q14, #9 yading@10: vshrn.i32 d19, q15, #9 yading@10: vadd.u16 q2, q2, q8 yading@10: vadd.u16 q3, q3, q9 yading@10: vrshrn.i16 d2, q2, #5 yading@10: vrshrn.i16 d3, q3, #5 yading@10: .endm yading@10: yading@10: /* void ff_rv40_weight_func_16_neon(uint8_t *dst, uint8_t *src1, uint8_t *src2, yading@10: int w1, int w2, int stride) */ yading@10: function ff_rv40_weight_func_16_neon, export=1 yading@10: ldr r12, [sp] yading@10: vmov d0, r3, r12 yading@10: ldr r12, [sp, #4] yading@10: mov r3, #16 yading@10: 1: yading@10: vld1.8 {q1}, [r1,:128], r12 yading@10: vld1.8 {q2}, [r2,:128], r12 yading@10: rv40_weight yading@10: vst1.8 {q1}, [r0,:128], r12 yading@10: subs r3, r3, #1 yading@10: bne 1b yading@10: bx lr yading@10: endfunc yading@10: yading@10: /* void ff_rv40_weight_func_8_neon(uint8_t *dst, uint8_t *src1, uint8_t *src2, yading@10: int w1, int w2, int stride) */ yading@10: function ff_rv40_weight_func_8_neon, export=1 yading@10: ldr r12, [sp] yading@10: vmov d0, r3, r12 yading@10: ldr r12, [sp, #4] yading@10: mov r3, #8 yading@10: 1: yading@10: vld1.8 {d2}, [r1,:64], r12 yading@10: vld1.8 {d3}, [r1,:64], r12 yading@10: vld1.8 {d4}, [r2,:64], r12 yading@10: vld1.8 {d5}, [r2,:64], r12 yading@10: rv40_weight yading@10: vst1.8 {d2}, [r0,:64], r12 yading@10: vst1.8 {d3}, [r0,:64], r12 yading@10: subs r3, r3, #2 yading@10: bne 1b yading@10: bx lr yading@10: endfunc yading@10: yading@10: function ff_rv40_h_loop_filter_strength_neon, export=1 yading@10: pkhbt r2, r3, r2, lsl #18 yading@10: yading@10: ldr r3, [r0] yading@10: ldr_dpre r12, r0, r1 yading@10: teq r3, r12 yading@10: beq 1f yading@10: yading@10: sub r0, r0, r1, lsl #1 yading@10: yading@10: vld1.32 {d4[]}, [r0,:32], r1 @ -3 yading@10: vld1.32 {d0[]}, [r0,:32], r1 @ -2 yading@10: vld1.32 {d4[1]}, [r0,:32], r1 @ -1 yading@10: vld1.32 {d5[]}, [r0,:32], r1 @ 0 yading@10: vld1.32 {d1[]}, [r0,:32], r1 @ 1 yading@10: vld1.32 {d5[0]}, [r0,:32], r1 @ 2 yading@10: yading@10: vpaddl.u8 q8, q0 @ -2, -2, -2, -2, 1, 1, 1, 1 yading@10: vpaddl.u8 q9, q2 @ -3, -3, -1, -1, 2, 2, 0, 0 yading@10: vdup.32 d30, r2 @ beta2, beta << 2 yading@10: vpadd.u16 d16, d16, d17 @ -2, -2, 1, 1 yading@10: vpadd.u16 d18, d18, d19 @ -3, -1, 2, 0 yading@10: vabd.u16 d16, d18, d16 yading@10: vclt.u16 d16, d16, d30 yading@10: yading@10: ldrd r2, r3, [sp, #4] yading@10: vmovl.u16 q12, d16 yading@10: vtrn.16 d16, d17 yading@10: vshr.u32 q12, q12, #15 yading@10: ldr r0, [sp] yading@10: vst1.32 {d24[1]}, [r2,:32] yading@10: vst1.32 {d25[1]}, [r3,:32] yading@10: yading@10: cmp r0, #0 yading@10: it eq yading@10: bxeq lr yading@10: yading@10: vand d18, d16, d17 yading@10: vtrn.32 d18, d19 yading@10: vand d18, d18, d19 yading@10: vmov.u16 r0, d18[0] yading@10: bx lr yading@10: 1: yading@10: ldrd r2, r3, [sp, #4] yading@10: mov r0, #0 yading@10: str r0, [r2] yading@10: str r0, [r3] yading@10: bx lr yading@10: endfunc yading@10: yading@10: function ff_rv40_v_loop_filter_strength_neon, export=1 yading@10: sub r0, r0, #3 yading@10: pkhbt r2, r3, r2, lsl #18 yading@10: yading@10: vld1.8 {d0}, [r0], r1 yading@10: vld1.8 {d1}, [r0], r1 yading@10: vld1.8 {d2}, [r0], r1 yading@10: vld1.8 {d3}, [r0], r1 yading@10: yading@10: vaddl.u8 q0, d0, d1 yading@10: vaddl.u8 q1, d2, d3 yading@10: vdup.32 q15, r2 yading@10: vadd.u16 q0, q0, q1 @ -3, -2, -1, 0, 1, 2 yading@10: vext.16 q1, q0, q0, #1 @ -2, -1, 0, 1, 2 yading@10: vabd.u16 q0, q1, q0 yading@10: vclt.u16 q0, q0, q15 yading@10: yading@10: ldrd r2, r3, [sp, #4] yading@10: vmovl.u16 q1, d0 yading@10: vext.16 d1, d0, d1, #3 yading@10: vshr.u32 q1, q1, #15 yading@10: ldr r0, [sp] yading@10: vst1.32 {d2[1]}, [r2,:32] yading@10: vst1.32 {d3[1]}, [r3,:32] yading@10: yading@10: cmp r0, #0 yading@10: it eq yading@10: bxeq lr yading@10: yading@10: vand d0, d0, d1 yading@10: vtrn.16 d0, d1 yading@10: vand d0, d0, d1 yading@10: vmov.u16 r0, d0[0] yading@10: bx lr yading@10: endfunc yading@10: yading@10: .macro rv40_weak_loop_filter yading@10: vdup.16 d30, r2 @ filter_p1 yading@10: vdup.16 d31, r3 @ filter_q1 yading@10: ldrd r2, r3, [sp] yading@10: vdup.16 d28, r2 @ alpha yading@10: vdup.16 d29, r3 @ beta yading@10: ldr r12, [sp, #8] yading@10: vdup.16 d25, r12 @ lim_p0q0 yading@10: ldrd r2, r3, [sp, #12] yading@10: vsubl.u8 q9, d5, d4 @ x, t yading@10: vabdl.u8 q8, d5, d4 @ x, abs(t) yading@10: vneg.s16 q15, q15 yading@10: vceq.i16 d16, d19, #0 @ !t yading@10: vshl.s16 d19, d19, #2 @ t << 2 yading@10: vmul.u16 d18, d17, d28 @ alpha * abs(t) yading@10: vand d24, d30, d31 @ filter_p1 & filter_q1 yading@10: vsubl.u8 q1, d0, d4 @ p1p2, p1p0 yading@10: vsubl.u8 q3, d1, d5 @ q1q2, q1q0 yading@10: vmov.i16 d22, #3 yading@10: vshr.u16 d18, d18, #7 yading@10: vadd.i16 d22, d22, d24 @ 3 - (filter_p1 & filter_q1) yading@10: vsubl.u8 q10, d0, d1 @ src[-2] - src[1] yading@10: vcle.u16 d18, d18, d22 yading@10: vand d20, d20, d24 yading@10: vneg.s16 d23, d25 @ -lim_p0q0 yading@10: vadd.s16 d19, d19, d20 yading@10: vbic d16, d18, d16 @ t && u <= 3 - (fp1 & fq1) yading@10: vtrn.32 d4, d5 @ -3, 2, -1, 0 yading@10: vrshr.s16 d19, d19, #3 yading@10: vmov d28, d29 @ beta yading@10: vswp d3, d6 @ q1q2, p1p0 yading@10: vmin.s16 d19, d19, d25 yading@10: vand d30, d30, d16 yading@10: vand d31, d31, d16 yading@10: vadd.s16 q10, q1, q3 @ p1p2 + p1p0, q1q2 + q1q0 yading@10: vmax.s16 d19, d19, d23 @ diff yading@10: vabs.s16 q1, q1 @ abs(p1p2), abs(q1q2) yading@10: vand d18, d19, d16 @ diff yading@10: vcle.u16 q1, q1, q14 yading@10: vneg.s16 d19, d18 @ -diff yading@10: vdup.16 d26, r3 @ lim_p1 yading@10: vaddw.u8 q2, q9, d5 @ src[-1]+diff, src[0]-diff yading@10: vhsub.s16 q11, q10, q9 yading@10: vand q1, q1, q15 yading@10: vqmovun.s16 d4, q2 @ -1, 0 yading@10: vand q9, q11, q1 yading@10: vdup.16 d27, r2 @ lim_q1 yading@10: vneg.s16 q9, q9 yading@10: vneg.s16 q14, q13 yading@10: vmin.s16 q9, q9, q13 yading@10: vtrn.32 d0, d1 @ -2, 1, -2, 1 yading@10: vmax.s16 q9, q9, q14 yading@10: vaddw.u8 q3, q9, d0 yading@10: vqmovun.s16 d5, q3 @ -2, 1 yading@10: .endm yading@10: yading@10: function ff_rv40_h_weak_loop_filter_neon, export=1 yading@10: sub r0, r0, r1, lsl #1 yading@10: sub r0, r0, r1 yading@10: yading@10: vld1.32 {d4[]}, [r0,:32], r1 yading@10: vld1.32 {d0[]}, [r0,:32], r1 yading@10: vld1.32 {d4[1]}, [r0,:32], r1 yading@10: vld1.32 {d5[]}, [r0,:32], r1 yading@10: vld1.32 {d1[]}, [r0,:32], r1 yading@10: vld1.32 {d5[0]}, [r0,:32] yading@10: yading@10: sub r0, r0, r1, lsl #2 yading@10: yading@10: rv40_weak_loop_filter yading@10: yading@10: vst1.32 {d5[0]}, [r0,:32], r1 yading@10: vst1.32 {d4[0]}, [r0,:32], r1 yading@10: vst1.32 {d4[1]}, [r0,:32], r1 yading@10: vst1.32 {d5[1]}, [r0,:32], r1 yading@10: yading@10: bx lr yading@10: endfunc yading@10: yading@10: function ff_rv40_v_weak_loop_filter_neon, export=1 yading@10: sub r12, r0, #3 yading@10: sub r0, r0, #2 yading@10: yading@10: vld1.8 {d4}, [r12], r1 yading@10: vld1.8 {d5}, [r12], r1 yading@10: vld1.8 {d2}, [r12], r1 yading@10: vld1.8 {d3}, [r12], r1 yading@10: yading@10: vtrn.16 q2, q1 yading@10: vtrn.8 d4, d5 yading@10: vtrn.8 d2, d3 yading@10: yading@10: vrev64.32 d5, d5 yading@10: vtrn.32 q2, q1 yading@10: vdup.32 d0, d3[0] yading@10: vdup.32 d1, d2[0] yading@10: yading@10: rv40_weak_loop_filter yading@10: yading@10: vtrn.32 q2, q3 yading@10: vswp d4, d5 yading@10: yading@10: vst4.8 {d4[0],d5[0],d6[0],d7[0]}, [r0], r1 yading@10: vst4.8 {d4[1],d5[1],d6[1],d7[1]}, [r0], r1 yading@10: vst4.8 {d4[2],d5[2],d6[2],d7[2]}, [r0], r1 yading@10: vst4.8 {d4[3],d5[3],d6[3],d7[3]}, [r0], r1 yading@10: yading@10: bx lr yading@10: endfunc