/*
* Copyright (c) 2011 Janne Grunau <janne-libav@jannau.net>
+ * Copyright (c) 2011 Mans Rullgard <mans@mansr.com>
*
* This file is part of Libav.
*
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
-#include "asm.S"
+#include "libavutil/arm/asm.S"
+#include "neon.S"
+
+.macro qpel_lowpass r0, r1, rc1, rc2, shift
+ vext.8 d25, \r0, \r1, #1 @ src[-1]
+ vext.8 d26, \r0, \r1, #4 @ src[ 2]
+ vext.8 d24, \r0, \r1, #5 @ src[ 3]
+ vaddl.u8 q9, d25, d26
+ vaddl.u8 q8, \r0, d24
+ vext.8 d27, \r0, \r1, #2 @ src[ 0]
+ vshl.s16 q12, q9, #2
+ vsub.s16 q8, q8, q9
+ vext.8 d28, \r0, \r1, #3 @ src[ 1]
+ vsub.s16 q8, q8, q12
+ vmlal.u8 q8, d27, \rc1
+ vmlal.u8 q8, d28, \rc2
+ vqrshrun.s16 \r0, q8, #\shift
+.endm
+
+.macro qpel_lowpass_x2 r0, r1, r2, r3, rc1, rc2, shift
+ vext.8 d25, \r0, \r1, #1 @ src[-1]
+ vext.8 d26, \r0, \r1, #4 @ src[ 2]
+ vext.8 d24, \r0, \r1, #5 @ src[ 3]
+ vaddl.u8 q9, d25, d26
+ vaddl.u8 q8, \r0, d24
+ vext.8 d29, \r0, \r1, #2 @ src[ 0]
+ vext.8 d28, \r0, \r1, #3 @ src[ 1]
+ vshl.s16 q10, q9, #2
+ vext.8 \r1, \r2, \r3, #1 @ src[-1]
+ vsub.s16 q8, q8, q9
+ vext.8 d22, \r2, \r3, #4 @ src[ 2]
+ vext.8 \r0, \r2, \r3, #5 @ src[ 3]
+ vaddl.u8 q13, \r1, d22
+ vaddl.u8 q12, \r2, \r0
+ vsub.s16 q8, q8, q10
+ vshl.s16 q9, q13, #2
+ vsub.s16 q12, q12, q13
+ vmlal.u8 q8, d29, \rc1
+ vmlal.u8 q8, d28, \rc2
+ vsub.s16 q12, q12, q9
+ vext.8 d26, \r2, \r3, #2 @ src[ 0]
+ vext.8 d27, \r2, \r3, #3 @ src[ 1]
+ vmlal.u8 q12, d26, \rc1
+ vmlal.u8 q12, d27, \rc2
+ vqrshrun.s16 \r0, q8, #\shift
+ vqrshrun.s16 \r2, q12, #\shift
+.endm
+
+.macro rv40_qpel8_h shift
+function put_rv40_qpel8_h_lp_packed_s\shift\()_neon
+1:
+ vld1.8 {q2}, [r1], r2
+ vld1.8 {q3}, [r1], r2
+ qpel_lowpass_x2 d4, d5, d6, d7, d0, d1, \shift
+ vst1.8 {d4}, [r12,:64]!
+ vst1.8 {d6}, [r12,:64]!
+ subs r3, r3, #2
+ bgt 1b
+ vld1.8 {q2}, [r1]
+ qpel_lowpass d4, d5, d0, d1, \shift
+ vst1.8 {d4}, [r12,:64]!
+ bx lr
+endfunc
+.endm
+
+.macro rv40_qpel8_v shift, type
+function \type\()_rv40_qpel8_v_lp_packed_s\shift\()_neon
+ vld1.64 {d2}, [r1,:64]!
+ vld1.64 {d3}, [r1,:64]!
+ vld1.64 {d4}, [r1,:64]!
+ vld1.64 {d5}, [r1,:64]!
+ vld1.64 {d6}, [r1,:64]!
+ vld1.64 {d7}, [r1,:64]!
+ vld1.64 {d8}, [r1,:64]!
+ vld1.64 {d9}, [r1,:64]!
+ vld1.64 {d10}, [r1,:64]!
+ vld1.64 {d11}, [r1,:64]!
+ vld1.64 {d12}, [r1,:64]!
+ vld1.64 {d13}, [r1,:64]!
+ vld1.64 {d14}, [r1,:64]!
+ transpose_8x8 d2, d3, d4, d5, d6, d7, d8, d9
+ transpose_8x8 d10, d11, d12, d13, d14, d15, d30, d31
+ qpel_lowpass_x2 d2, d10, d3, d11, d0, d1, \shift
+ qpel_lowpass_x2 d4, d12, d5, d13, d0, d1, \shift
+ qpel_lowpass_x2 d6, d14, d7, d15, d0, d1, \shift
+ qpel_lowpass_x2 d8, d30, d9, d31, d0, d1, \shift
+ transpose_8x8 d2, d3, d4, d5, d6, d7, d8, d9
+ .ifc \type,avg
+ vld1.64 d12, [r0,:64], r2
+ vld1.64 d13, [r0,:64], r2
+ vld1.64 d14, [r0,:64], r2
+ vld1.64 d15, [r0,:64], r2
+ vld1.64 d16, [r0,:64], r2
+ vld1.64 d17, [r0,:64], r2
+ vld1.64 d18, [r0,:64], r2
+ vld1.64 d19, [r0,:64], r2
+ sub r0, r0, r2, lsl #3
+ vrhadd.u8 q1, q1, q6
+ vrhadd.u8 q2, q2, q7
+ vrhadd.u8 q3, q3, q8
+ vrhadd.u8 q4, q4, q9
+ .endif
+ vst1.64 d2, [r0,:64], r2
+ vst1.64 d3, [r0,:64], r2
+ vst1.64 d4, [r0,:64], r2
+ vst1.64 d5, [r0,:64], r2
+ vst1.64 d6, [r0,:64], r2
+ vst1.64 d7, [r0,:64], r2
+ vst1.64 d8, [r0,:64], r2
+ vst1.64 d9, [r0,:64], r2
+ bx lr
+endfunc
+.endm
+
+ rv40_qpel8_h 5
+ rv40_qpel8_h 6
+
+.macro rv40_qpel type
+function \type\()_rv40_qpel8_h_lowpass_neon
+ .ifc \type,avg
+ mov r12, r0
+ .endif
+1:
+ vld1.8 {q2}, [r1], r2
+ vld1.8 {q3}, [r1], r2
+ qpel_lowpass_x2 d4, d5, d6, d7, d0, d1, 6
+ .ifc \type,avg
+ vld1.8 {d3}, [r12,:64], r2
+ vld1.8 {d16}, [r12,:64], r2
+ vrhadd.u8 d4, d4, d3
+ vrhadd.u8 d6, d6, d16
+ .endif
+ vst1.8 {d4}, [r0,:64], r2
+ vst1.8 {d6}, [r0,:64], r2
+ subs r3, r3, #2
+ bgt 1b
+ bx lr
+endfunc
+
+function \type\()_rv40_qpel8_v_lowpass_neon
+ vld1.64 {d2}, [r1], r2
+ vld1.64 {d3}, [r1], r2
+ vld1.64 {d4}, [r1], r2
+ vld1.64 {d5}, [r1], r2
+ vld1.64 {d6}, [r1], r2
+ vld1.64 {d7}, [r1], r2
+ vld1.64 {d8}, [r1], r2
+ vld1.64 {d9}, [r1], r2
+ vld1.64 {d10}, [r1], r2
+ vld1.64 {d11}, [r1], r2
+ vld1.64 {d12}, [r1], r2
+ vld1.64 {d13}, [r1], r2
+ vld1.64 {d14}, [r1]
+ transpose_8x8 d2, d3, d4, d5, d6, d7, d8, d9
+ transpose_8x8 d10, d11, d12, d13, d14, d15, d30, d31
+ qpel_lowpass_x2 d2, d10, d3, d11, d0, d1, 6
+ qpel_lowpass_x2 d4, d12, d5, d13, d0, d1, 6
+ qpel_lowpass_x2 d6, d14, d7, d15, d0, d1, 6
+ qpel_lowpass_x2 d8, d30, d9, d31, d0, d1, 6
+ transpose_8x8 d2, d3, d4, d5, d6, d7, d8, d9
+ .ifc \type,avg
+ vld1.64 d12, [r0,:64], r2
+ vld1.64 d13, [r0,:64], r2
+ vld1.64 d14, [r0,:64], r2
+ vld1.64 d15, [r0,:64], r2
+ vld1.64 d16, [r0,:64], r2
+ vld1.64 d17, [r0,:64], r2
+ vld1.64 d18, [r0,:64], r2
+ vld1.64 d19, [r0,:64], r2
+ sub r0, r0, r2, lsl #3
+ vrhadd.u8 q1, q1, q6
+ vrhadd.u8 q2, q2, q7
+ vrhadd.u8 q3, q3, q8
+ vrhadd.u8 q4, q4, q9
+ .endif
+ vst1.64 d2, [r0,:64], r2
+ vst1.64 d3, [r0,:64], r2
+ vst1.64 d4, [r0,:64], r2
+ vst1.64 d5, [r0,:64], r2
+ vst1.64 d6, [r0,:64], r2
+ vst1.64 d7, [r0,:64], r2
+ vst1.64 d8, [r0,:64], r2
+ vst1.64 d9, [r0,:64], r2
+ bx lr
+endfunc
+
+ rv40_qpel8_v 5, \type
+ rv40_qpel8_v 6, \type
+
+function ff_\type\()_rv40_qpel8_mc10_neon, export=1
+ sub r1, r1, #2
+ mov r3, #8
+ vmov.i8 d0, #52
+ vmov.i8 d1, #20
+ b \type\()_rv40_qpel8_h_lowpass_neon
+endfunc
+
+function ff_\type\()_rv40_qpel8_mc30_neon, export=1
+ sub r1, r1, #2
+ mov r3, #8
+ vmov.i8 d0, #20
+ vmov.i8 d1, #52
+ b \type\()_rv40_qpel8_h_lowpass_neon
+endfunc
+
+function ff_\type\()_rv40_qpel8_mc01_neon, export=1
+ push {r4, lr}
+ vpush {d8-d15}
+ sub r1, r1, r2, lsl #1
+ vmov.i8 d0, #52
+ vmov.i8 d1, #20
+ bl \type\()_rv40_qpel8_v_lowpass_neon
+ vpop {d8-d15}
+ pop {r4, pc}
+endfunc
+
+function ff_\type\()_rv40_qpel8_mc11_neon, export=1
+ push {r4, lr}
+ vpush {d8-d15}
+ sub sp, sp, #14*8
+ add r12, sp, #7
+ bic r12, r12, #7
+ sub r1, r1, r2, lsl #1
+ sub r1, r1, #2
+ mov r3, #12
+ vmov.i8 d0, #52
+ vmov.i8 d1, #20
+ bl put_rv40_qpel8_h_lp_packed_s6_neon
+ add r1, sp, #7
+ bic r1, r1, #7
+ bl \type\()_rv40_qpel8_v_lp_packed_s6_neon
+ add sp, sp, #14*8
+ vpop {d8-d15}
+ pop {r4, pc}
+endfunc
+
+function ff_\type\()_rv40_qpel8_mc21_neon, export=1
+ push {r4, lr}
+ vpush {d8-d15}
+ sub sp, sp, #14*8
+ add r12, sp, #7
+ bic r12, r12, #7
+ sub r1, r1, r2, lsl #1
+ sub r1, r1, #2
+ mov r3, #12
+ vmov.i8 d0, #20
+ vmov.i8 d1, #20
+ bl put_rv40_qpel8_h_lp_packed_s5_neon
+ add r1, sp, #7
+ bic r1, r1, #7
+ vmov.i8 d0, #52
+ bl \type\()_rv40_qpel8_v_lp_packed_s6_neon
+ add sp, sp, #14*8
+ vpop {d8-d15}
+ pop {r4, pc}
+endfunc
+
+function ff_\type\()_rv40_qpel8_mc31_neon, export=1
+ push {r4, lr}
+ vpush {d8-d15}
+ sub sp, sp, #14*8
+ add r12, sp, #7
+ bic r12, r12, #7
+ sub r1, r1, r2, lsl #1
+ sub r1, r1, #2
+ mov r3, #12
+ vmov.i8 d0, #20
+ vmov.i8 d1, #52
+ bl put_rv40_qpel8_h_lp_packed_s6_neon
+ add r1, sp, #7
+ bic r1, r1, #7
+ vswp d0, d1
+ bl \type\()_rv40_qpel8_v_lp_packed_s6_neon
+ add sp, sp, #14*8
+ vpop {d8-d15}
+ pop {r4, pc}
+endfunc
+
+function ff_\type\()_rv40_qpel8_mc12_neon, export=1
+ push {r4, lr}
+ vpush {d8-d15}
+ sub sp, sp, #14*8
+ add r12, sp, #7
+ bic r12, r12, #7
+ sub r1, r1, r2, lsl #1
+ sub r1, r1, #2
+ mov r3, #12
+ vmov.i8 d0, #52
+ vmov.i8 d1, #20
+ bl put_rv40_qpel8_h_lp_packed_s6_neon
+ add r1, sp, #7
+ bic r1, r1, #7
+ vmov.i8 d0, #20
+ bl \type\()_rv40_qpel8_v_lp_packed_s5_neon
+ add sp, sp, #14*8
+ vpop {d8-d15}
+ pop {r4, pc}
+endfunc
+
+function ff_\type\()_rv40_qpel8_mc22_neon, export=1
+ push {r4, lr}
+ vpush {d8-d15}
+ sub sp, sp, #14*8
+ add r12, sp, #7
+ bic r12, r12, #7
+ sub r1, r1, r2, lsl #1
+ sub r1, r1, #2
+ mov r3, #12
+ vmov.i8 d0, #20
+ vmov.i8 d1, #20
+ bl put_rv40_qpel8_h_lp_packed_s5_neon
+ add r1, sp, #7
+ bic r1, r1, #7
+ bl \type\()_rv40_qpel8_v_lp_packed_s5_neon
+ add sp, sp, #14*8
+ vpop {d8-d15}
+ pop {r4, pc}
+endfunc
+
+function ff_\type\()_rv40_qpel8_mc32_neon, export=1
+ push {r4, lr}
+ vpush {d8-d15}
+ sub sp, sp, #14*8
+ add r12, sp, #7
+ bic r12, r12, #7
+ sub r1, r1, r2, lsl #1
+ sub r1, r1, #2
+ mov r3, #12
+ vmov.i8 d0, #20
+ vmov.i8 d1, #52
+ bl put_rv40_qpel8_h_lp_packed_s6_neon
+ add r1, sp, #7
+ bic r1, r1, #7
+ vmov.i8 d1, #20
+ bl \type\()_rv40_qpel8_v_lp_packed_s5_neon
+ add sp, sp, #14*8
+ vpop {d8-d15}
+ pop {r4, pc}
+endfunc
+
+function ff_\type\()_rv40_qpel8_mc03_neon, export=1
+ push {r4, lr}
+ vpush {d8-d15}
+ sub r1, r1, r2, lsl #1
+ vmov.i8 d0, #20
+ vmov.i8 d1, #52
+ bl \type\()_rv40_qpel8_v_lowpass_neon
+ vpop {d8-d15}
+ pop {r4, pc}
+endfunc
+
+function ff_\type\()_rv40_qpel8_mc33_neon, export=1
+ mov r3, #8
+ b X(ff_\type\()_pixels8_xy2_neon)
+endfunc
+
+function ff_\type\()_rv40_qpel8_mc13_neon, export=1
+ push {r4, lr}
+ vpush {d8-d15}
+ sub sp, sp, #14*8
+ add r12, sp, #7
+ bic r12, r12, #7
+ sub r1, r1, r2, lsl #1
+ sub r1, r1, #2
+ mov r3, #12
+ vmov.i8 d0, #52
+ vmov.i8 d1, #20
+ bl put_rv40_qpel8_h_lp_packed_s6_neon
+ add r1, sp, #7
+ bic r1, r1, #7
+ vswp d0, d1
+ bl \type\()_rv40_qpel8_v_lp_packed_s6_neon
+ add sp, sp, #14*8
+ vpop {d8-d15}
+ pop {r4, pc}
+endfunc
+
+function ff_\type\()_rv40_qpel8_mc23_neon, export=1
+ push {r4, lr}
+ vpush {d8-d15}
+ sub sp, sp, #14*8
+ add r12, sp, #7
+ bic r12, r12, #7
+ sub r1, r1, r2, lsl #1
+ sub r1, r1, #2
+ mov r3, #12
+ vmov.i8 d0, #20
+ vmov.i8 d1, #20
+ bl put_rv40_qpel8_h_lp_packed_s5_neon
+ add r1, sp, #7
+ bic r1, r1, #7
+ vmov.i8 d1, #52
+ bl \type\()_rv40_qpel8_v_lp_packed_s6_neon
+ add sp, sp, #14*8
+ vpop {d8-d15}
+ pop {r4, pc}
+endfunc
+
+function ff_\type\()_rv40_qpel16_mc10_neon, export=1
+ vmov.i8 d0, #52
+ vmov.i8 d1, #20
+.L\type\()_rv40_qpel16_h:
+ push {r1, lr}
+ sub r1, r1, #2
+ mov r3, #16
+ bl \type\()_rv40_qpel8_h_lowpass_neon
+ pop {r1, lr}
+ sub r0, r0, r2, lsl #4
+ add r0, r0, #8
+ add r1, r1, #6
+ mov r3, #16
+ b \type\()_rv40_qpel8_h_lowpass_neon
+endfunc
+
+function ff_\type\()_rv40_qpel16_mc30_neon, export=1
+ vmov.i8 d0, #20
+ vmov.i8 d1, #52
+ b .L\type\()_rv40_qpel16_h
+endfunc
+
+function ff_\type\()_rv40_qpel16_mc01_neon, export=1
+ vmov.i8 d0, #52
+ vmov.i8 d1, #20
+.L\type\()_rv40_qpel16_v:
+ sub r1, r1, r2, lsl #1
+ push {r1, lr}
+ vpush {d8-d15}
+ bl \type\()_rv40_qpel8_v_lowpass_neon
+ sub r1, r1, r2, lsl #2
+ bl \type\()_rv40_qpel8_v_lowpass_neon
+ ldr r1, [sp, #64]
+ sub r0, r0, r2, lsl #4
+ add r0, r0, #8
+ add r1, r1, #8
+ bl \type\()_rv40_qpel8_v_lowpass_neon
+ sub r1, r1, r2, lsl #2
+ bl \type\()_rv40_qpel8_v_lowpass_neon
+ vpop {d8-d15}
+ pop {r1, pc}
+endfunc
+
+function ff_\type\()_rv40_qpel16_mc11_neon, export=1
+ sub r1, r1, r2, lsl #1
+ sub r1, r1, #2
+ push {r1, lr}
+ vpush {d8-d15}
+ sub sp, sp, #44*8
+ add r12, sp, #7
+ bic r12, r12, #7
+ mov r3, #20
+ vmov.i8 d0, #52
+ vmov.i8 d1, #20
+ bl put_rv40_qpel8_h_lp_packed_s6_neon
+ ldr r1, [sp, #416]
+ add r1, r1, #8
+ mov r3, #20
+ bl put_rv40_qpel8_h_lp_packed_s6_neon
+.L\type\()_rv40_qpel16_v_s6:
+ add r1, sp, #7
+ bic r1, r1, #7
+ bl \type\()_rv40_qpel8_v_lp_packed_s6_neon
+ sub r1, r1, #40
+ bl \type\()_rv40_qpel8_v_lp_packed_s6_neon
+ sub r0, r0, r2, lsl #4
+ add r0, r0, #8
+ bl \type\()_rv40_qpel8_v_lp_packed_s6_neon
+ sub r1, r1, #40
+ bl \type\()_rv40_qpel8_v_lp_packed_s6_neon
+ add sp, sp, #44*8
+ vpop {d8-d15}
+ pop {r1, pc}
+endfunc
+
+function ff_\type\()_rv40_qpel16_mc21_neon, export=1
+ sub r1, r1, r2, lsl #1
+ sub r1, r1, #2
+ push {r1, lr}
+ vpush {d8-d15}
+ sub sp, sp, #44*8
+ add r12, sp, #7
+ bic r12, r12, #7
+ mov r3, #20
+ vmov.i8 d0, #20
+ vmov.i8 d1, #20
+ bl put_rv40_qpel8_h_lp_packed_s5_neon
+ ldr r1, [sp, #416]
+ add r1, r1, #8
+ mov r3, #20
+ bl put_rv40_qpel8_h_lp_packed_s5_neon
+ vmov.i8 d0, #52
+ b .L\type\()_rv40_qpel16_v_s6
+endfunc
+
+function ff_\type\()_rv40_qpel16_mc31_neon, export=1
+ sub r1, r1, r2, lsl #1
+ sub r1, r1, #2
+ push {r1, lr}
+ vpush {d8-d15}
+ sub sp, sp, #44*8
+ add r12, sp, #7
+ bic r12, r12, #7
+ mov r3, #20
+ vmov.i8 d0, #20
+ vmov.i8 d1, #52
+ bl put_rv40_qpel8_h_lp_packed_s6_neon
+ ldr r1, [sp, #416]
+ add r1, r1, #8
+ mov r3, #20
+ bl put_rv40_qpel8_h_lp_packed_s6_neon
+ vswp d0, d1
+ b .L\type\()_rv40_qpel16_v_s6
+endfunc
+
+function ff_\type\()_rv40_qpel16_mc12_neon, export=1
+ sub r1, r1, r2, lsl #1
+ sub r1, r1, #2
+ push {r1, lr}
+ vpush {d8-d15}
+ sub sp, sp, #44*8
+ add r12, sp, #7
+ bic r12, r12, #7
+ mov r3, #20
+ vmov.i8 d0, #52
+ vmov.i8 d1, #20
+ bl put_rv40_qpel8_h_lp_packed_s6_neon
+ ldr r1, [sp, #416]
+ add r1, r1, #8
+ mov r3, #20
+ bl put_rv40_qpel8_h_lp_packed_s6_neon
+ vmov.i8 d0, #20
+.L\type\()_rv40_qpel16_v_s5:
+ add r1, sp, #7
+ bic r1, r1, #7
+ bl \type\()_rv40_qpel8_v_lp_packed_s5_neon
+ sub r1, r1, #40
+ bl \type\()_rv40_qpel8_v_lp_packed_s5_neon
+ sub r0, r0, r2, lsl #4
+ add r0, r0, #8
+ bl \type\()_rv40_qpel8_v_lp_packed_s5_neon
+ sub r1, r1, #40
+ bl \type\()_rv40_qpel8_v_lp_packed_s5_neon
+ add sp, sp, #44*8
+ vpop {d8-d15}
+ pop {r1, pc}
+endfunc
+
+function ff_\type\()_rv40_qpel16_mc22_neon, export=1
+ sub r1, r1, r2, lsl #1
+ sub r1, r1, #2
+ push {r1, lr}
+ vpush {d8-d15}
+ sub sp, sp, #44*8
+ add r12, sp, #7
+ bic r12, r12, #7
+ mov r3, #20
+ vmov.i8 d0, #20
+ vmov.i8 d1, #20
+ bl put_rv40_qpel8_h_lp_packed_s5_neon
+ ldr r1, [sp, #416]
+ add r1, r1, #8
+ mov r3, #20
+ bl put_rv40_qpel8_h_lp_packed_s5_neon
+ b .L\type\()_rv40_qpel16_v_s5
+endfunc
+
+function ff_\type\()_rv40_qpel16_mc32_neon, export=1
+ sub r1, r1, r2, lsl #1
+ sub r1, r1, #2
+ push {r1, lr}
+ vpush {d8-d15}
+ sub sp, sp, #44*8
+ add r12, sp, #7
+ bic r12, r12, #7
+ mov r3, #20
+ vmov.i8 d0, #20
+ vmov.i8 d1, #52
+ bl put_rv40_qpel8_h_lp_packed_s6_neon
+ ldr r1, [sp, #416]
+ add r1, r1, #8
+ mov r3, #20
+ bl put_rv40_qpel8_h_lp_packed_s6_neon
+ vmov.i8 d1, #20
+ b .L\type\()_rv40_qpel16_v_s5
+endfunc
+
+function ff_\type\()_rv40_qpel16_mc03_neon, export=1
+ vmov.i8 d0, #20
+ vmov.i8 d1, #52
+ b .L\type\()_rv40_qpel16_v
+endfunc
+
+function ff_\type\()_rv40_qpel16_mc13_neon, export=1
+ sub r1, r1, r2, lsl #1
+ sub r1, r1, #2
+ push {r1, lr}
+ vpush {d8-d15}
+ sub sp, sp, #44*8
+ add r12, sp, #7
+ bic r12, r12, #7
+ mov r3, #20
+ vmov.i8 d0, #52
+ vmov.i8 d1, #20
+ bl put_rv40_qpel8_h_lp_packed_s6_neon
+ ldr r1, [sp, #416]
+ add r1, r1, #8
+ mov r3, #20
+ bl put_rv40_qpel8_h_lp_packed_s6_neon
+ vswp d0, d1
+ b .L\type\()_rv40_qpel16_v_s6
+endfunc
+
+function ff_\type\()_rv40_qpel16_mc23_neon, export=1
+ sub r1, r1, r2, lsl #1
+ sub r1, r1, #2
+ push {r1, lr}
+ vpush {d8-d15}
+ sub sp, sp, #44*8
+ add r12, sp, #7
+ bic r12, r12, #7
+ mov r3, #20
+ vmov.i8 d0, #20
+ vmov.i8 d1, #20
+ bl put_rv40_qpel8_h_lp_packed_s5_neon
+ ldr r1, [sp, #416]
+ add r1, r1, #8
+ mov r3, #20
+ bl put_rv40_qpel8_h_lp_packed_s5_neon
+ vmov.i8 d1, #52
+ b .L\type\()_rv40_qpel16_v_s6
+endfunc
+
+function ff_\type\()_rv40_qpel16_mc33_neon, export=1
+ mov r3, #16
+ b X(ff_\type\()_pixels16_xy2_neon)
+endfunc
+.endm
+
+ rv40_qpel put
+ rv40_qpel avg
.macro rv40_weight
vmovl.u8 q8, d2
bne 1b
bx lr
endfunc
+
+function ff_rv40_h_loop_filter_strength_neon, export=1
+ pkhbt r2, r3, r2, lsl #18
+
+ ldr r3, [r0]
+ ldr_dpre r12, r0, r1
+ teq r3, r12
+ beq 1f
+
+ sub r0, r0, r1, lsl #1
+
+ vld1.32 {d4[]}, [r0,:32], r1 @ -3
+ vld1.32 {d0[]}, [r0,:32], r1 @ -2
+ vld1.32 {d4[1]}, [r0,:32], r1 @ -1
+ vld1.32 {d5[]}, [r0,:32], r1 @ 0
+ vld1.32 {d1[]}, [r0,:32], r1 @ 1
+ vld1.32 {d5[0]}, [r0,:32], r1 @ 2
+
+ vpaddl.u8 q8, q0 @ -2, -2, -2, -2, 1, 1, 1, 1
+ vpaddl.u8 q9, q2 @ -3, -3, -1, -1, 2, 2, 0, 0
+ vdup.32 d30, r2 @ beta2, beta << 2
+ vpadd.u16 d16, d16, d17 @ -2, -2, 1, 1
+ vpadd.u16 d18, d18, d19 @ -3, -1, 2, 0
+ vabd.u16 d16, d18, d16
+ vclt.u16 d16, d16, d30
+
+ ldrd r2, r3, [sp, #4]
+ vmovl.u16 q12, d16
+ vtrn.16 d16, d17
+ vshr.u32 q12, q12, #15
+ ldr r0, [sp]
+ vst1.32 {d24[1]}, [r2,:32]
+ vst1.32 {d25[1]}, [r3,:32]
+
+ cmp r0, #0
+ it eq
+ bxeq lr
+
+ vand d18, d16, d17
+ vtrn.32 d18, d19
+ vand d18, d18, d19
+ vmov.u16 r0, d18[0]
+ bx lr
+1:
+ ldrd r2, r3, [sp, #4]
+ mov r0, #0
+ str r0, [r2]
+ str r0, [r3]
+ bx lr
+endfunc
+
+function ff_rv40_v_loop_filter_strength_neon, export=1
+ sub r0, r0, #3
+ pkhbt r2, r3, r2, lsl #18
+
+ vld1.8 {d0}, [r0], r1
+ vld1.8 {d1}, [r0], r1
+ vld1.8 {d2}, [r0], r1
+ vld1.8 {d3}, [r0], r1
+
+ vaddl.u8 q0, d0, d1
+ vaddl.u8 q1, d2, d3
+ vdup.32 q15, r2
+ vadd.u16 q0, q0, q1 @ -3, -2, -1, 0, 1, 2
+ vext.16 q1, q0, q0, #1 @ -2, -1, 0, 1, 2
+ vabd.u16 q0, q1, q0
+ vclt.u16 q0, q0, q15
+
+ ldrd r2, r3, [sp, #4]
+ vmovl.u16 q1, d0
+ vext.16 d1, d0, d1, #3
+ vshr.u32 q1, q1, #15
+ ldr r0, [sp]
+ vst1.32 {d2[1]}, [r2,:32]
+ vst1.32 {d3[1]}, [r3,:32]
+
+ cmp r0, #0
+ it eq
+ bxeq lr
+
+ vand d0, d0, d1
+ vtrn.16 d0, d1
+ vand d0, d0, d1
+ vmov.u16 r0, d0[0]
+ bx lr
+endfunc
+
+.macro rv40_weak_loop_filter
+ vdup.16 d30, r2 @ filter_p1
+ vdup.16 d31, r3 @ filter_q1
+ ldrd r2, r3, [sp]
+ vdup.16 d28, r2 @ alpha
+ vdup.16 d29, r3 @ beta
+ ldr r12, [sp, #8]
+ vdup.16 d25, r12 @ lim_p0q0
+ ldrd r2, r3, [sp, #12]
+ vsubl.u8 q9, d5, d4 @ x, t
+ vabdl.u8 q8, d5, d4 @ x, abs(t)
+ vneg.s16 q15, q15
+ vceq.i16 d16, d19, #0 @ !t
+ vshl.s16 d19, d19, #2 @ t << 2
+ vmul.u16 d18, d17, d28 @ alpha * abs(t)
+ vand d24, d30, d31 @ filter_p1 & filter_q1
+ vsubl.u8 q1, d0, d4 @ p1p2, p1p0
+ vsubl.u8 q3, d1, d5 @ q1q2, q1q0
+ vmov.i16 d22, #3
+ vshr.u16 d18, d18, #7
+ vadd.i16 d22, d22, d24 @ 3 - (filter_p1 & filter_q1)
+ vsubl.u8 q10, d0, d1 @ src[-2] - src[1]
+ vcle.u16 d18, d18, d22
+ vand d20, d20, d24
+ vneg.s16 d23, d25 @ -lim_p0q0
+ vadd.s16 d19, d19, d20
+ vbic d16, d18, d16 @ t && u <= 3 - (fp1 & fq1)
+ vtrn.32 d4, d5 @ -3, 2, -1, 0
+ vrshr.s16 d19, d19, #3
+ vmov d28, d29 @ beta
+ vswp d3, d6 @ q1q2, p1p0
+ vmin.s16 d19, d19, d25
+ vand d30, d30, d16
+ vand d31, d31, d16
+ vadd.s16 q10, q1, q3 @ p1p2 + p1p0, q1q2 + q1q0
+ vmax.s16 d19, d19, d23 @ diff
+ vabs.s16 q1, q1 @ abs(p1p2), abs(q1q2)
+ vand d18, d19, d16 @ diff
+ vcle.u16 q1, q1, q14
+ vneg.s16 d19, d18 @ -diff
+ vdup.16 d26, r3 @ lim_p1
+ vaddw.u8 q2, q9, d5 @ src[-1]+diff, src[0]-diff
+ vhsub.s16 q11, q10, q9
+ vand q1, q1, q15
+ vqmovun.s16 d4, q2 @ -1, 0
+ vand q9, q11, q1
+ vdup.16 d27, r2 @ lim_q1
+ vneg.s16 q9, q9
+ vneg.s16 q14, q13
+ vmin.s16 q9, q9, q13
+ vtrn.32 d0, d1 @ -2, 1, -2, 1
+ vmax.s16 q9, q9, q14
+ vaddw.u8 q3, q9, d0
+ vqmovun.s16 d5, q3 @ -2, 1
+.endm
+
+function ff_rv40_h_weak_loop_filter_neon, export=1
+ sub r0, r0, r1, lsl #1
+ sub r0, r0, r1
+
+ vld1.32 {d4[]}, [r0,:32], r1
+ vld1.32 {d0[]}, [r0,:32], r1
+ vld1.32 {d4[1]}, [r0,:32], r1
+ vld1.32 {d5[]}, [r0,:32], r1
+ vld1.32 {d1[]}, [r0,:32], r1
+ vld1.32 {d5[0]}, [r0,:32]
+
+ sub r0, r0, r1, lsl #2
+
+ rv40_weak_loop_filter
+
+ vst1.32 {d5[0]}, [r0,:32], r1
+ vst1.32 {d4[0]}, [r0,:32], r1
+ vst1.32 {d4[1]}, [r0,:32], r1
+ vst1.32 {d5[1]}, [r0,:32], r1
+
+ bx lr
+endfunc
+
+function ff_rv40_v_weak_loop_filter_neon, export=1
+ sub r12, r0, #3
+ sub r0, r0, #2
+
+ vld1.8 {d4}, [r12], r1
+ vld1.8 {d5}, [r12], r1
+ vld1.8 {d2}, [r12], r1
+ vld1.8 {d3}, [r12], r1
+
+ vtrn.16 q2, q1
+ vtrn.8 d4, d5
+ vtrn.8 d2, d3
+
+ vrev64.32 d5, d5
+ vtrn.32 q2, q1
+ vdup.32 d0, d3[0]
+ vdup.32 d1, d2[0]
+
+ rv40_weak_loop_filter
+
+ vtrn.32 q2, q3
+ vswp d4, d5
+
+ vst4.8 {d4[0],d5[0],d6[0],d7[0]}, [r0], r1
+ vst4.8 {d4[1],d5[1],d6[1],d7[1]}, [r0], r1
+ vst4.8 {d4[2],d5[2],d6[2],d7[2]}, [r0], r1
+ vst4.8 {d4[3],d5[3],d6[3],d7[3]}, [r0], r1
+
+ bx lr
+endfunc