pjs/media/libtheora/lib/arm/armloop.s

677 строки
23 KiB
ArmAsm

;********************************************************************
;* *
;* THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE. *
;* USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS *
;* GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
;* IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING. *
;* *
;* THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2010 *
;* by the Xiph.Org Foundation and contributors http://www.xiph.org/ *
;* *
;********************************************************************
; Original implementation:
; Copyright (C) 2009 Robin Watts for Pinknoise Productions Ltd
; last mod: $Id: armloop.s 17481 2010-10-03 22:49:42Z tterribe $
;********************************************************************
AREA |.text|, CODE, READONLY
GET armopts.s
EXPORT oc_loop_filter_frag_rows_arm
; Which bit this is depends on the order of packing within a bitfield.
; Hopefully that doesn't change among any of the relevant compilers.
OC_FRAG_CODED_FLAG * 1
; Vanilla ARM v4 version
loop_filter_h_arm PROC
; r0 = unsigned char *_pix
; r1 = int _ystride
; r2 = int *_bv
; preserves r0-r3
STMFD r13!,{r3-r6,r14}
MOV r14,#8
MOV r6, #255
lfh_arm_lp
LDRB r3, [r0, #-2] ; r3 = _pix[0]
LDRB r12,[r0, #1] ; r12= _pix[3]
LDRB r4, [r0, #-1] ; r4 = _pix[1]
LDRB r5, [r0] ; r5 = _pix[2]
SUB r3, r3, r12 ; r3 = _pix[0]-_pix[3]+4
ADD r3, r3, #4
SUB r12,r5, r4 ; r12= _pix[2]-_pix[1]
ADD r12,r12,r12,LSL #1 ; r12= 3*(_pix[2]-_pix[1])
ADD r12,r12,r3 ; r12= _pix[0]-_pix[3]+3*(_pix[2]-_pix[1])+4
MOV r12,r12,ASR #3
LDRSB r12,[r2, r12]
; Stall (2 on Xscale)
ADDS r4, r4, r12
CMPGT r6, r4
EORLT r4, r6, r4, ASR #32
SUBS r5, r5, r12
CMPGT r6, r5
EORLT r5, r6, r5, ASR #32
STRB r4, [r0, #-1]
STRB r5, [r0], r1
SUBS r14,r14,#1
BGT lfh_arm_lp
SUB r0, r0, r1, LSL #3
LDMFD r13!,{r3-r6,PC}
ENDP
loop_filter_v_arm PROC
; r0 = unsigned char *_pix
; r1 = int _ystride
; r2 = int *_bv
; preserves r0-r3
STMFD r13!,{r3-r6,r14}
MOV r14,#8
MOV r6, #255
lfv_arm_lp
LDRB r3, [r0, -r1, LSL #1] ; r3 = _pix[0]
LDRB r12,[r0, r1] ; r12= _pix[3]
LDRB r4, [r0, -r1] ; r4 = _pix[1]
LDRB r5, [r0] ; r5 = _pix[2]
SUB r3, r3, r12 ; r3 = _pix[0]-_pix[3]+4
ADD r3, r3, #4
SUB r12,r5, r4 ; r12= _pix[2]-_pix[1]
ADD r12,r12,r12,LSL #1 ; r12= 3*(_pix[2]-_pix[1])
ADD r12,r12,r3 ; r12= _pix[0]-_pix[3]+3*(_pix[2]-_pix[1])+4
MOV r12,r12,ASR #3
LDRSB r12,[r2, r12]
; Stall (2 on Xscale)
ADDS r4, r4, r12
CMPGT r6, r4
EORLT r4, r6, r4, ASR #32
SUBS r5, r5, r12
CMPGT r6, r5
EORLT r5, r6, r5, ASR #32
STRB r4, [r0, -r1]
STRB r5, [r0], #1
SUBS r14,r14,#1
BGT lfv_arm_lp
SUB r0, r0, #8
LDMFD r13!,{r3-r6,PC}
ENDP
oc_loop_filter_frag_rows_arm PROC
; r0 = _ref_frame_data
; r1 = _ystride
; r2 = _bv
; r3 = _frags
; r4 = _fragi0
; r5 = _fragi0_end
; r6 = _fragi_top
; r7 = _fragi_bot
; r8 = _frag_buf_offs
; r9 = _nhfrags
MOV r12,r13
STMFD r13!,{r0,r4-r11,r14}
LDMFD r12,{r4-r9}
ADD r2, r2, #127 ; _bv += 127
CMP r4, r5 ; if(_fragi0>=_fragi0_end)
BGE oslffri_arm_end ; bail
SUBS r9, r9, #1 ; r9 = _nhfrags-1 if (r9<=0)
BLE oslffri_arm_end ; bail
ADD r3, r3, r4, LSL #2 ; r3 = &_frags[fragi]
ADD r8, r8, r4, LSL #2 ; r8 = &_frag_buf_offs[fragi]
SUB r7, r7, r9 ; _fragi_bot -= _nhfrags;
oslffri_arm_lp1
MOV r10,r4 ; r10= fragi = _fragi0
ADD r11,r4, r9 ; r11= fragi_end-1=fragi+_nhfrags-1
oslffri_arm_lp2
LDR r14,[r3], #4 ; r14= _frags[fragi] _frags++
LDR r0, [r13] ; r0 = _ref_frame_data
LDR r12,[r8], #4 ; r12= _frag_buf_offs[fragi] _frag_buf_offs++
TST r14,#OC_FRAG_CODED_FLAG
BEQ oslffri_arm_uncoded
CMP r10,r4 ; if (fragi>_fragi0)
ADD r0, r0, r12 ; r0 = _ref_frame_data + _frag_buf_offs[fragi]
BLGT loop_filter_h_arm
CMP r4, r6 ; if (_fragi0>_fragi_top)
BLGT loop_filter_v_arm
CMP r10,r11 ; if(fragi+1<fragi_end)===(fragi<fragi_end-1)
LDRLT r12,[r3] ; r12 = _frags[fragi+1]
ADD r0, r0, #8
ADD r10,r10,#1 ; r10 = fragi+1;
ANDLT r12,r12,#OC_FRAG_CODED_FLAG
CMPLT r12,#OC_FRAG_CODED_FLAG ; && _frags[fragi+1].coded==0
BLLT loop_filter_h_arm
CMP r10,r7 ; if (fragi<_fragi_bot)
LDRLT r12,[r3, r9, LSL #2] ; r12 = _frags[fragi+1+_nhfrags-1]
SUB r0, r0, #8
ADD r0, r0, r1, LSL #3
ANDLT r12,r12,#OC_FRAG_CODED_FLAG
CMPLT r12,#OC_FRAG_CODED_FLAG
BLLT loop_filter_v_arm
CMP r10,r11 ; while(fragi<=fragi_end-1)
BLE oslffri_arm_lp2
MOV r4, r10 ; r4 = fragi0 += _nhfrags
CMP r4, r5
BLT oslffri_arm_lp1
oslffri_arm_end
LDMFD r13!,{r0,r4-r11,PC}
oslffri_arm_uncoded
ADD r10,r10,#1
CMP r10,r11
BLE oslffri_arm_lp2
MOV r4, r10 ; r4 = _fragi0 += _nhfrags
CMP r4, r5
BLT oslffri_arm_lp1
LDMFD r13!,{r0,r4-r11,PC}
ENDP
[ OC_ARM_ASM_MEDIA
EXPORT oc_loop_filter_init_v6
EXPORT oc_loop_filter_frag_rows_v6
oc_loop_filter_init_v6 PROC
; r0 = _bv
; r1 = _flimit (=L from the spec)
MVN r1, r1, LSL #1 ; r1 = <0xFFFFFF|255-2*L>
AND r1, r1, #255 ; r1 = ll=r1&0xFF
ORR r1, r1, r1, LSL #8 ; r1 = <ll|ll>
PKHBT r1, r1, r1, LSL #16 ; r1 = <ll|ll|ll|ll>
STR r1, [r0]
MOV PC,r14
ENDP
; We could use the same strategy as the v filter below, but that would require
; 40 instructions to load the data and transpose it into columns and another
; 32 to write out the results at the end, plus the 52 instructions to do the
; filtering itself.
; This is slightly less, and less code, even assuming we could have shared the
; 52 instructions in the middle with the other function.
; It executes slightly fewer instructions than the ARMv6 approach David Conrad
; proposed for FFmpeg, but not by much:
; http://lists.mplayerhq.hu/pipermail/ffmpeg-devel/2010-February/083141.html
; His is a lot less code, though, because it only does two rows at once instead
; of four.
loop_filter_h_v6 PROC
; r0 = unsigned char *_pix
; r1 = int _ystride
; r2 = int _ll
; preserves r0-r3
STMFD r13!,{r4-r11,r14}
LDR r12,=0x10003
BL loop_filter_h_core_v6
ADD r0, r0, r1, LSL #2
BL loop_filter_h_core_v6
SUB r0, r0, r1, LSL #2
LDMFD r13!,{r4-r11,PC}
ENDP
loop_filter_h_core_v6 PROC
; r0 = unsigned char *_pix
; r1 = int _ystride
; r2 = int _ll
; r12= 0x10003
; Preserves r0-r3, r12; Clobbers r4-r11.
LDR r4,[r0, #-2]! ; r4 = <p3|p2|p1|p0>
; Single issue
LDR r5,[r0, r1]! ; r5 = <q3|q2|q1|q0>
UXTB16 r6, r4, ROR #16 ; r6 = <p0|p2>
UXTB16 r4, r4, ROR #8 ; r4 = <p3|p1>
UXTB16 r7, r5, ROR #16 ; r7 = <q0|q2>
UXTB16 r5, r5, ROR #8 ; r5 = <q3|q1>
PKHBT r8, r4, r5, LSL #16 ; r8 = <__|q1|__|p1>
PKHBT r9, r6, r7, LSL #16 ; r9 = <__|q2|__|p2>
SSUB16 r6, r4, r6 ; r6 = <p3-p0|p1-p2>
SMLAD r6, r6, r12,r12 ; r6 = <????|(p3-p0)+3*(p1-p2)+3>
SSUB16 r7, r5, r7 ; r7 = <q3-q0|q1-q2>
SMLAD r7, r7, r12,r12 ; r7 = <????|(q0-q3)+3*(q2-q1)+4>
LDR r4,[r0, r1]! ; r4 = <r3|r2|r1|r0>
MOV r6, r6, ASR #3 ; r6 = <??????|(p3-p0)+3*(p1-p2)+3>>3>
LDR r5,[r0, r1]! ; r5 = <s3|s2|s1|s0>
PKHBT r11,r6, r7, LSL #13 ; r11= <??|-R_q|??|-R_p>
UXTB16 r6, r4, ROR #16 ; r6 = <r0|r2>
UXTB16 r11,r11 ; r11= <__|-R_q|__|-R_p>
UXTB16 r4, r4, ROR #8 ; r4 = <r3|r1>
UXTB16 r7, r5, ROR #16 ; r7 = <s0|s2>
PKHBT r10,r6, r7, LSL #16 ; r10= <__|s2|__|r2>
SSUB16 r6, r4, r6 ; r6 = <r3-r0|r1-r2>
UXTB16 r5, r5, ROR #8 ; r5 = <s3|s1>
SMLAD r6, r6, r12,r12 ; r6 = <????|(r3-r0)+3*(r2-r1)+3>
SSUB16 r7, r5, r7 ; r7 = <r3-r0|r1-r2>
SMLAD r7, r7, r12,r12 ; r7 = <????|(s0-s3)+3*(s2-s1)+4>
ORR r9, r9, r10, LSL #8 ; r9 = <s2|q2|r2|p2>
MOV r6, r6, ASR #3 ; r6 = <??????|(r0-r3)+3*(r2-r1)+4>>3>
PKHBT r10,r4, r5, LSL #16 ; r10= <__|s1|__|r1>
PKHBT r6, r6, r7, LSL #13 ; r6 = <??|-R_s|??|-R_r>
ORR r8, r8, r10, LSL #8 ; r8 = <s1|q1|r1|p1>
UXTB16 r6, r6 ; r6 = <__|-R_s|__|-R_r>
MOV r10,#0
ORR r6, r11,r6, LSL #8 ; r6 = <-R_s|-R_q|-R_r|-R_p>
; Single issue
; There's no min, max or abs instruction.
; SSUB8 and SEL will work for abs, and we can do all the rest with
; unsigned saturated adds, which means the GE flags are still all
; set when we're done computing lflim(abs(R_i),L).
; This allows us to both add and subtract, and split the results by
; the original sign of R_i.
SSUB8 r7, r10,r6
; Single issue
SEL r7, r7, r6 ; r7 = abs(R_i)
; Single issue
UQADD8 r4, r7, r2 ; r4 = 255-max(2*L-abs(R_i),0)
; Single issue
UQADD8 r7, r7, r4
; Single issue
UQSUB8 r7, r7, r4 ; r7 = min(abs(R_i),max(2*L-abs(R_i),0))
; Single issue
UQSUB8 r4, r8, r7
UQADD8 r5, r9, r7
UQADD8 r8, r8, r7
UQSUB8 r9, r9, r7
SEL r8, r8, r4 ; r8 = p1+lflim(R_i,L)
SEL r9, r9, r5 ; r9 = p2-lflim(R_i,L)
MOV r5, r9, LSR #24 ; r5 = s2
STRB r5, [r0,#2]!
MOV r4, r8, LSR #24 ; r4 = s1
STRB r4, [r0,#-1]
MOV r5, r9, LSR #8 ; r5 = r2
STRB r5, [r0,-r1]!
MOV r4, r8, LSR #8 ; r4 = r1
STRB r4, [r0,#-1]
MOV r5, r9, LSR #16 ; r5 = q2
STRB r5, [r0,-r1]!
MOV r4, r8, LSR #16 ; r4 = q1
STRB r4, [r0,#-1]
; Single issue
STRB r9, [r0,-r1]!
; Single issue
STRB r8, [r0,#-1]
MOV PC,r14
ENDP
; This uses the same strategy as the MMXEXT version for x86, except that UHADD8
; computes (a+b>>1) instead of (a+b+1>>1) like PAVGB.
; This works just as well, with the following procedure for computing the
; filter value, f:
; u = ~UHADD8(p1,~p2);
; v = UHADD8(~p1,p2);
; m = v-u;
; a = m^UHADD8(m^p0,m^~p3);
; f = UHADD8(UHADD8(a,u1),v1);
; where f = 127+R, with R in [-127,128] defined as in the spec.
; This is exactly the same amount of arithmetic as the version that uses PAVGB
; as the basic operator.
; It executes about 2/3 the number of instructions of David Conrad's approach,
; but requires more code, because it does all eight columns at once, instead
; of four at a time.
loop_filter_v_v6 PROC
; r0 = unsigned char *_pix
; r1 = int _ystride
; r2 = int _ll
; preserves r0-r11
STMFD r13!,{r4-r11,r14}
LDRD r6, [r0, -r1]! ; r7, r6 = <p5|p1>
LDRD r4, [r0, -r1] ; r5, r4 = <p4|p0>
LDRD r8, [r0, r1]! ; r9, r8 = <p6|p2>
MVN r14,r6 ; r14= ~p1
LDRD r10,[r0, r1] ; r11,r10= <p7|p3>
; Filter the first four columns.
MVN r12,r8 ; r12= ~p2
UHADD8 r14,r14,r8 ; r14= v1=~p1+p2>>1
UHADD8 r12,r12,r6 ; r12= p1+~p2>>1
MVN r10, r10 ; r10=~p3
MVN r12,r12 ; r12= u1=~p1+p2+1>>1
SSUB8 r14,r14,r12 ; r14= m1=v1-u1
; Single issue
EOR r4, r4, r14 ; r4 = m1^p0
EOR r10,r10,r14 ; r10= m1^~p3
UHADD8 r4, r4, r10 ; r4 = (m1^p0)+(m1^~p3)>>1
; Single issue
EOR r4, r4, r14 ; r4 = a1=m1^((m1^p0)+(m1^~p3)>>1)
SADD8 r14,r14,r12 ; r14= v1=m1+u1
UHADD8 r4, r4, r12 ; r4 = a1+u1>>1
MVN r12,r9 ; r12= ~p6
UHADD8 r4, r4, r14 ; r4 = f1=(a1+u1>>1)+v1>>1
; Filter the second four columns.
MVN r14,r7 ; r14= ~p5
UHADD8 r12,r12,r7 ; r12= p5+~p6>>1
UHADD8 r14,r14,r9 ; r14= v2=~p5+p6>>1
MVN r12,r12 ; r12= u2=~p5+p6+1>>1
MVN r11,r11 ; r11=~p7
SSUB8 r10,r14,r12 ; r10= m2=v2-u2
; Single issue
EOR r5, r5, r10 ; r5 = m2^p4
EOR r11,r11,r10 ; r11= m2^~p7
UHADD8 r5, r5, r11 ; r5 = (m2^p4)+(m2^~p7)>>1
; Single issue
EOR r5, r5, r10 ; r5 = a2=m2^((m2^p4)+(m2^~p7)>>1)
; Single issue
UHADD8 r5, r5, r12 ; r5 = a2+u2>>1
LDR r12,=0x7F7F7F7F ; r12 = {127}x4
UHADD8 r5, r5, r14 ; r5 = f2=(a2+u2>>1)+v2>>1
; Now split f[i] by sign.
; There's no min or max instruction.
; We could use SSUB8 and SEL, but this is just as many instructions and
; dual issues more (for v7 without NEON).
UQSUB8 r10,r4, r12 ; r10= R_i>0?R_i:0
UQSUB8 r4, r12,r4 ; r4 = R_i<0?-R_i:0
UQADD8 r11,r10,r2 ; r11= 255-max(2*L-abs(R_i<0),0)
UQADD8 r14,r4, r2 ; r14= 255-max(2*L-abs(R_i>0),0)
UQADD8 r10,r10,r11
UQADD8 r4, r4, r14
UQSUB8 r10,r10,r11 ; r10= min(abs(R_i<0),max(2*L-abs(R_i<0),0))
UQSUB8 r4, r4, r14 ; r4 = min(abs(R_i>0),max(2*L-abs(R_i>0),0))
UQSUB8 r11,r5, r12 ; r11= R_i>0?R_i:0
UQADD8 r6, r6, r10
UQSUB8 r8, r8, r10
UQSUB8 r5, r12,r5 ; r5 = R_i<0?-R_i:0
UQSUB8 r6, r6, r4 ; r6 = p1+lflim(R_i,L)
UQADD8 r8, r8, r4 ; r8 = p2-lflim(R_i,L)
UQADD8 r10,r11,r2 ; r10= 255-max(2*L-abs(R_i<0),0)
UQADD8 r14,r5, r2 ; r14= 255-max(2*L-abs(R_i>0),0)
UQADD8 r11,r11,r10
UQADD8 r5, r5, r14
UQSUB8 r11,r11,r10 ; r11= min(abs(R_i<0),max(2*L-abs(R_i<0),0))
UQSUB8 r5, r5, r14 ; r5 = min(abs(R_i>0),max(2*L-abs(R_i>0),0))
UQADD8 r7, r7, r11
UQSUB8 r9, r9, r11
UQSUB8 r7, r7, r5 ; r7 = p5+lflim(R_i,L)
STRD r6, [r0, -r1] ; [p5:p1] = [r7: r6]
UQADD8 r9, r9, r5 ; r9 = p6-lflim(R_i,L)
STRD r8, [r0] ; [p6:p2] = [r9: r8]
LDMFD r13!,{r4-r11,PC}
ENDP
oc_loop_filter_frag_rows_v6 PROC
; r0 = _ref_frame_data
; r1 = _ystride
; r2 = _bv
; r3 = _frags
; r4 = _fragi0
; r5 = _fragi0_end
; r6 = _fragi_top
; r7 = _fragi_bot
; r8 = _frag_buf_offs
; r9 = _nhfrags
MOV r12,r13
STMFD r13!,{r0,r4-r11,r14}
LDMFD r12,{r4-r9}
LDR r2, [r2] ; ll = *(int *)_bv
CMP r4, r5 ; if(_fragi0>=_fragi0_end)
BGE oslffri_v6_end ; bail
SUBS r9, r9, #1 ; r9 = _nhfrags-1 if (r9<=0)
BLE oslffri_v6_end ; bail
ADD r3, r3, r4, LSL #2 ; r3 = &_frags[fragi]
ADD r8, r8, r4, LSL #2 ; r8 = &_frag_buf_offs[fragi]
SUB r7, r7, r9 ; _fragi_bot -= _nhfrags;
oslffri_v6_lp1
MOV r10,r4 ; r10= fragi = _fragi0
ADD r11,r4, r9 ; r11= fragi_end-1=fragi+_nhfrags-1
oslffri_v6_lp2
LDR r14,[r3], #4 ; r14= _frags[fragi] _frags++
LDR r0, [r13] ; r0 = _ref_frame_data
LDR r12,[r8], #4 ; r12= _frag_buf_offs[fragi] _frag_buf_offs++
TST r14,#OC_FRAG_CODED_FLAG
BEQ oslffri_v6_uncoded
CMP r10,r4 ; if (fragi>_fragi0)
ADD r0, r0, r12 ; r0 = _ref_frame_data + _frag_buf_offs[fragi]
BLGT loop_filter_h_v6
CMP r4, r6 ; if (fragi0>_fragi_top)
BLGT loop_filter_v_v6
CMP r10,r11 ; if(fragi+1<fragi_end)===(fragi<fragi_end-1)
LDRLT r12,[r3] ; r12 = _frags[fragi+1]
ADD r0, r0, #8
ADD r10,r10,#1 ; r10 = fragi+1;
ANDLT r12,r12,#OC_FRAG_CODED_FLAG
CMPLT r12,#OC_FRAG_CODED_FLAG ; && _frags[fragi+1].coded==0
BLLT loop_filter_h_v6
CMP r10,r7 ; if (fragi<_fragi_bot)
LDRLT r12,[r3, r9, LSL #2] ; r12 = _frags[fragi+1+_nhfrags-1]
SUB r0, r0, #8
ADD r0, r0, r1, LSL #3
ANDLT r12,r12,#OC_FRAG_CODED_FLAG
CMPLT r12,#OC_FRAG_CODED_FLAG
BLLT loop_filter_v_v6
CMP r10,r11 ; while(fragi<=fragi_end-1)
BLE oslffri_v6_lp2
MOV r4, r10 ; r4 = fragi0 += nhfrags
CMP r4, r5
BLT oslffri_v6_lp1
oslffri_v6_end
LDMFD r13!,{r0,r4-r11,PC}
oslffri_v6_uncoded
ADD r10,r10,#1
CMP r10,r11
BLE oslffri_v6_lp2
MOV r4, r10 ; r4 = fragi0 += nhfrags
CMP r4, r5
BLT oslffri_v6_lp1
LDMFD r13!,{r0,r4-r11,PC}
ENDP
]
[ OC_ARM_ASM_NEON
EXPORT oc_loop_filter_init_neon
EXPORT oc_loop_filter_frag_rows_neon
oc_loop_filter_init_neon PROC
; r0 = _bv
; r1 = _flimit (=L from the spec)
MOV r1, r1, LSL #1 ; r1 = 2*L
VDUP.S16 Q15, r1 ; Q15= 2L in U16s
VST1.64 {D30,D31}, [r0@128]
MOV PC,r14
ENDP
loop_filter_h_neon PROC
; r0 = unsigned char *_pix
; r1 = int _ystride
; r2 = int *_bv
; preserves r0-r3
; We assume Q15= 2*L in U16s
; My best guesses at cycle counts (and latency)--vvv
SUB r12,r0, #2
; Doing a 2-element structure load saves doing two VTRN's below, at the
; cost of using two more slower single-lane loads vs. the faster
; all-lane loads.
; It's less code this way, though, and benches a hair faster, but it
; leaves D2 and D4 swapped.
VLD2.16 {D0[],D2[]}, [r12], r1 ; D0 = ____________1100 2,1
; D2 = ____________3322
VLD2.16 {D4[],D6[]}, [r12], r1 ; D4 = ____________5544 2,1
; D6 = ____________7766
VLD2.16 {D0[1],D2[1]},[r12], r1 ; D0 = ________99881100 3,1
; D2 = ________BBAA3322
VLD2.16 {D4[1],D6[1]},[r12], r1 ; D4 = ________DDCC5544 3,1
; D6 = ________FFEE7766
VLD2.16 {D0[2],D2[2]},[r12], r1 ; D0 = ____GGHH99881100 3,1
; D2 = ____JJIIBBAA3322
VLD2.16 {D4[2],D6[2]},[r12], r1 ; D4 = ____KKLLDDCC5544 3,1
; D6 = ____NNMMFFEE7766
VLD2.16 {D0[3],D2[3]},[r12], r1 ; D0 = PPOOGGHH99881100 3,1
; D2 = RRQQJJIIBBAA3322
VLD2.16 {D4[3],D6[3]},[r12], r1 ; D4 = TTSSKKLLDDCC5544 3,1
; D6 = VVUUNNMMFFEE7766
VTRN.8 D0, D4 ; D0 = SSOOKKGGCC884400 D4 = TTPPLLHHDD995511 1,1
VTRN.8 D2, D6 ; D2 = UUQQMMIIEEAA6622 D6 = VVRRNNJJFFBB7733 1,1
VSUBL.U8 Q0, D0, D6 ; Q0 = 00 - 33 in S16s 1,3
VSUBL.U8 Q8, D2, D4 ; Q8 = 22 - 11 in S16s 1,3
ADD r12,r0, #8
VADD.S16 Q0, Q0, Q8 ; 1,3
PLD [r12]
VADD.S16 Q0, Q0, Q8 ; 1,3
PLD [r12,r1]
VADD.S16 Q0, Q0, Q8 ; Q0 = [0-3]+3*[2-1] 1,3
PLD [r12,r1, LSL #1]
VRSHR.S16 Q0, Q0, #3 ; Q0 = f = ([0-3]+3*[2-1]+4)>>3 1,4
ADD r12,r12,r1, LSL #2
; We want to do
; f = CLAMP(MIN(-2L-f,0), f, MAX(2L-f,0))
; = ((f >= 0) ? MIN( f ,MAX(2L- f ,0)) : MAX( f , MIN(-2L- f ,0)))
; = ((f >= 0) ? MIN(|f|,MAX(2L-|f|,0)) : MAX(-|f|, MIN(-2L+|f|,0)))
; = ((f >= 0) ? MIN(|f|,MAX(2L-|f|,0)) :-MIN( |f|,-MIN(-2L+|f|,0)))
; = ((f >= 0) ? MIN(|f|,MAX(2L-|f|,0)) :-MIN( |f|, MAX( 2L-|f|,0)))
; So we've reduced the left and right hand terms to be the same, except
; for a negation.
; Stall x3
VABS.S16 Q9, Q0 ; Q9 = |f| in U16s 1,4
PLD [r12,-r1]
VSHR.S16 Q0, Q0, #15 ; Q0 = -1 or 0 according to sign 1,3
PLD [r12]
VQSUB.U16 Q10,Q15,Q9 ; Q10= MAX(2L-|f|,0) in U16s 1,4
PLD [r12,r1]
VMOVL.U8 Q1, D2 ; Q2 = __UU__QQ__MM__II__EE__AA__66__22 2,3
PLD [r12,r1,LSL #1]
VMIN.U16 Q9, Q10,Q9 ; Q9 = MIN(|f|,MAX(2L-|f|)) 1,4
ADD r12,r12,r1, LSL #2
; Now we need to correct for the sign of f.
; For negative elements of Q0, we want to subtract the appropriate
; element of Q9. For positive elements we want to add them. No NEON
; instruction exists to do this, so we need to negate the negative
; elements, and we can then just add them. a-b = a-(1+!b) = a-1+!b
VADD.S16 Q9, Q9, Q0 ; 1,3
PLD [r12,-r1]
VEOR.S16 Q9, Q9, Q0 ; Q9 = real value of f 1,3
; Bah. No VRSBW.U8
; Stall (just 1 as Q9 not needed to second pipeline stage. I think.)
VADDW.U8 Q2, Q9, D4 ; Q1 = xxTTxxPPxxLLxxHHxxDDxx99xx55xx11 1,3
VSUB.S16 Q1, Q1, Q9 ; Q2 = xxUUxxQQxxMMxxIIxxEExxAAxx66xx22 1,3
VQMOVUN.S16 D4, Q2 ; D4 = TTPPLLHHDD995511 1,1
VQMOVUN.S16 D2, Q1 ; D2 = UUQQMMIIEEAA6622 1,1
SUB r12,r0, #1
VTRN.8 D4, D2 ; D4 = QQPPIIHHAA992211 D2 = MMLLEEDD6655 1,1
VST1.16 {D4[0]}, [r12], r1
VST1.16 {D2[0]}, [r12], r1
VST1.16 {D4[1]}, [r12], r1
VST1.16 {D2[1]}, [r12], r1
VST1.16 {D4[2]}, [r12], r1
VST1.16 {D2[2]}, [r12], r1
VST1.16 {D4[3]}, [r12], r1
VST1.16 {D2[3]}, [r12], r1
MOV PC,r14
ENDP
loop_filter_v_neon PROC
; r0 = unsigned char *_pix
; r1 = int _ystride
; r2 = int *_bv
; preserves r0-r3
; We assume Q15= 2*L in U16s
; My best guesses at cycle counts (and latency)--vvv
SUB r12,r0, r1, LSL #1
VLD1.64 {D0}, [r12@64], r1 ; D0 = SSOOKKGGCC884400 2,1
VLD1.64 {D2}, [r12@64], r1 ; D2 = TTPPLLHHDD995511 2,1
VLD1.64 {D4}, [r12@64], r1 ; D4 = UUQQMMIIEEAA6622 2,1
VLD1.64 {D6}, [r12@64] ; D6 = VVRRNNJJFFBB7733 2,1
VSUBL.U8 Q8, D4, D2 ; Q8 = 22 - 11 in S16s 1,3
VSUBL.U8 Q0, D0, D6 ; Q0 = 00 - 33 in S16s 1,3
ADD r12, #8
VADD.S16 Q0, Q0, Q8 ; 1,3
PLD [r12]
VADD.S16 Q0, Q0, Q8 ; 1,3
PLD [r12,r1]
VADD.S16 Q0, Q0, Q8 ; Q0 = [0-3]+3*[2-1] 1,3
SUB r12, r0, r1
VRSHR.S16 Q0, Q0, #3 ; Q0 = f = ([0-3]+3*[2-1]+4)>>3 1,4
; We want to do
; f = CLAMP(MIN(-2L-f,0), f, MAX(2L-f,0))
; = ((f >= 0) ? MIN( f ,MAX(2L- f ,0)) : MAX( f , MIN(-2L- f ,0)))
; = ((f >= 0) ? MIN(|f|,MAX(2L-|f|,0)) : MAX(-|f|, MIN(-2L+|f|,0)))
; = ((f >= 0) ? MIN(|f|,MAX(2L-|f|,0)) :-MIN( |f|,-MIN(-2L+|f|,0)))
; = ((f >= 0) ? MIN(|f|,MAX(2L-|f|,0)) :-MIN( |f|, MAX( 2L-|f|,0)))
; So we've reduced the left and right hand terms to be the same, except
; for a negation.
; Stall x3
VABS.S16 Q9, Q0 ; Q9 = |f| in U16s 1,4
VSHR.S16 Q0, Q0, #15 ; Q0 = -1 or 0 according to sign 1,3
; Stall x2
VQSUB.U16 Q10,Q15,Q9 ; Q10= MAX(2L-|f|,0) in U16s 1,4
VMOVL.U8 Q2, D4 ; Q2 = __UU__QQ__MM__II__EE__AA__66__22 2,3
; Stall x2
VMIN.U16 Q9, Q10,Q9 ; Q9 = MIN(|f|,MAX(2L-|f|)) 1,4
; Now we need to correct for the sign of f.
; For negative elements of Q0, we want to subtract the appropriate
; element of Q9. For positive elements we want to add them. No NEON
; instruction exists to do this, so we need to negate the negative
; elements, and we can then just add them. a-b = a-(1+!b) = a-1+!b
; Stall x3
VADD.S16 Q9, Q9, Q0 ; 1,3
; Stall x2
VEOR.S16 Q9, Q9, Q0 ; Q9 = real value of f 1,3
; Bah. No VRSBW.U8
; Stall (just 1 as Q9 not needed to second pipeline stage. I think.)
VADDW.U8 Q1, Q9, D2 ; Q1 = xxTTxxPPxxLLxxHHxxDDxx99xx55xx11 1,3
VSUB.S16 Q2, Q2, Q9 ; Q2 = xxUUxxQQxxMMxxIIxxEExxAAxx66xx22 1,3
VQMOVUN.S16 D2, Q1 ; D2 = TTPPLLHHDD995511 1,1
VQMOVUN.S16 D4, Q2 ; D4 = UUQQMMIIEEAA6622 1,1
VST1.64 {D2}, [r12@64], r1
VST1.64 {D4}, [r12@64], r1
MOV PC,r14
ENDP
oc_loop_filter_frag_rows_neon PROC
; r0 = _ref_frame_data
; r1 = _ystride
; r2 = _bv
; r3 = _frags
; r4 = _fragi0
; r5 = _fragi0_end
; r6 = _fragi_top
; r7 = _fragi_bot
; r8 = _frag_buf_offs
; r9 = _nhfrags
MOV r12,r13
STMFD r13!,{r0,r4-r11,r14}
LDMFD r12,{r4-r9}
CMP r4, r5 ; if(_fragi0>=_fragi0_end)
BGE oslffri_neon_end; bail
SUBS r9, r9, #1 ; r9 = _nhfrags-1 if (r9<=0)
BLE oslffri_neon_end ; bail
VLD1.64 {D30,D31}, [r2@128] ; Q15= 2L in U16s
ADD r3, r3, r4, LSL #2 ; r3 = &_frags[fragi]
ADD r8, r8, r4, LSL #2 ; r8 = &_frag_buf_offs[fragi]
SUB r7, r7, r9 ; _fragi_bot -= _nhfrags;
oslffri_neon_lp1
MOV r10,r4 ; r10= fragi = _fragi0
ADD r11,r4, r9 ; r11= fragi_end-1=fragi+_nhfrags-1
oslffri_neon_lp2
LDR r14,[r3], #4 ; r14= _frags[fragi] _frags++
LDR r0, [r13] ; r0 = _ref_frame_data
LDR r12,[r8], #4 ; r12= _frag_buf_offs[fragi] _frag_buf_offs++
TST r14,#OC_FRAG_CODED_FLAG
BEQ oslffri_neon_uncoded
CMP r10,r4 ; if (fragi>_fragi0)
ADD r0, r0, r12 ; r0 = _ref_frame_data + _frag_buf_offs[fragi]
BLGT loop_filter_h_neon
CMP r4, r6 ; if (_fragi0>_fragi_top)
BLGT loop_filter_v_neon
CMP r10,r11 ; if(fragi+1<fragi_end)===(fragi<fragi_end-1)
LDRLT r12,[r3] ; r12 = _frags[fragi+1]
ADD r0, r0, #8
ADD r10,r10,#1 ; r10 = fragi+1;
ANDLT r12,r12,#OC_FRAG_CODED_FLAG
CMPLT r12,#OC_FRAG_CODED_FLAG ; && _frags[fragi+1].coded==0
BLLT loop_filter_h_neon
CMP r10,r7 ; if (fragi<_fragi_bot)
LDRLT r12,[r3, r9, LSL #2] ; r12 = _frags[fragi+1+_nhfrags-1]
SUB r0, r0, #8
ADD r0, r0, r1, LSL #3
ANDLT r12,r12,#OC_FRAG_CODED_FLAG
CMPLT r12,#OC_FRAG_CODED_FLAG
BLLT loop_filter_v_neon
CMP r10,r11 ; while(fragi<=fragi_end-1)
BLE oslffri_neon_lp2
MOV r4, r10 ; r4 = _fragi0 += _nhfrags
CMP r4, r5
BLT oslffri_neon_lp1
oslffri_neon_end
LDMFD r13!,{r0,r4-r11,PC}
oslffri_neon_uncoded
ADD r10,r10,#1
CMP r10,r11
BLE oslffri_neon_lp2
MOV r4, r10 ; r4 = _fragi0 += _nhfrags
CMP r4, r5
BLT oslffri_neon_lp1
LDMFD r13!,{r0,r4-r11,PC}
ENDP
]
END