Source code
Revision control
Copy as Markdown
Other Tools
/*
* Copyright © 2018, VideoLAN and dav1d authors
* Copyright © 2020, Martin Storsjo
* All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions are met:
*
* 1. Redistributions of source code must retain the above copyright notice, this
* list of conditions and the following disclaimer.
*
* 2. Redistributions in binary form must reproduce the above copyright notice,
* this list of conditions and the following disclaimer in the documentation
* and/or other materials provided with the distribution.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
* ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
* WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
* ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
* (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
* LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
* ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
* SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/
#include "src/arm/asm.S"
#include "util.S"
const right_ext_mask_buf
.byte 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00
.byte 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00
.byte 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00
.byte 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00
right_ext_mask:
.byte 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff
.byte 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff
.byte 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff
.byte 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff
endconst
// void dav1d_wiener_filter_h_16bpc_neon(int16_t *dst, const pixel (*left)[4],
// const pixel *src, ptrdiff_t stride,
// const int16_t fh[7], const intptr_t w,
// int h, enum LrEdgeFlags edges,
// const int bitdepth_max);
function wiener_filter_h_16bpc_neon, export=1
push {r4-r11,lr}
vpush {q4-q7}
ldrd r4, r5, [sp, #100]
ldrd r6, r7, [sp, #108]
ldr r8, [sp, #116] // bitdepth_max
vld1.16 {q0}, [r4, :128]
clz r8, r8
vmov.i32 q14, #1
sub r9, r8, #38 // -(bitdepth + 6)
sub r8, r8, #25 // -round_bits_h
neg r9, r9 // bitdepth + 6
vdup.32 q1, r9
vdup.32 q13, r8 // -round_bits_h
vmov.i16 q15, #8192
vshl.u32 q14, q14, q1 // 1 << (bitdepth + 6)
mov r8, r5
// Calculate mid_stride
add r10, r5, #7
bic r10, r10, #7
lsl r10, r10, #1
// Set up pointers for reading/writing alternate rows
add r12, r0, r10
lsl r10, r10, #1
add lr, r2, r3
lsl r3, r3, #1
// Subtract the aligned width from mid_stride
add r11, r5, #7
bic r11, r11, #7
sub r10, r10, r11, lsl #1
// Subtract the number of pixels read from the source stride
add r11, r11, #8
sub r3, r3, r11, lsl #1
// Set up the src pointers to include the left edge, for LR_HAVE_LEFT, left == NULL
tst r7, #1 // LR_HAVE_LEFT
beq 2f
// LR_HAVE_LEFT
cmp r1, #0
bne 0f
// left == NULL
sub r2, r2, #6
sub lr, lr, #6
b 1f
0: // LR_HAVE_LEFT, left != NULL
2: // !LR_HAVE_LEFT, increase the stride.
// For this case we don't read the left 3 pixels from the src pointer,
// but shift it as if we had done that.
add r3, r3, #6
1: // Loop vertically
vld1.16 {q2, q3}, [r2]!
vld1.16 {q4, q5}, [lr]!
tst r7, #1 // LR_HAVE_LEFT
beq 0f
cmp r1, #0
beq 2f
// LR_HAVE_LEFT, left != NULL
vld1.16 {d3}, [r1]!
// Move r2/lr back to account for the last 3 pixels we loaded earlier,
// which we'll shift out.
sub r2, r2, #6
sub lr, lr, #6
vld1.16 {d13}, [r1]!
vext.8 q3, q2, q3, #10
vext.8 q2, q1, q2, #10
vext.8 q5, q4, q5, #10
vext.8 q4, q6, q4, #10
b 2f
0:
// !LR_HAVE_LEFT, fill q1 with the leftmost pixel
// and shift q2/q3 to have 3x the first pixel at the front.
vdup.16 q1, d4[0]
vdup.16 q6, d8[0]
// Move r2 back to account for the last 3 pixels we loaded before,
// which we shifted out.
sub r2, r2, #6
sub lr, lr, #6
vext.8 q3, q2, q3, #10
vext.8 q2, q1, q2, #10
vext.8 q5, q4, q5, #10
vext.8 q4, q6, q4, #10
2:
tst r7, #2 // LR_HAVE_RIGHT
bne 4f
// If we'll need to pad the right edge, load that pixel to pad with
// here since we can find it pretty easily from here.
sub r9, r5, #14
lsl r9, r9, #1
ldrh r11, [r2, r9]
ldrh r9, [lr, r9]
// Fill q11/q12 with the right padding pixel
vdup.16 q11, r11
vdup.16 q12, r9
3: // !LR_HAVE_RIGHT
// Check whether we need to pad the right edge
cmp r5, #11
bge 4f // If w >= 11, all used input pixels are valid
// 1 <= w < 11, w+3 pixels valid in q2-q3. For w=9 or w=10,
// this ends up called again; it's not strictly needed in those
// cases (we pad enough here), but keeping the code as simple as possible.
// Insert padding in q2/3.h[w+3] onwards; fuse the +3 (*2) into the
// buffer pointer.
movrel_local r4, right_ext_mask, -6
sub r4, r4, r5, lsl #1
vld1.8 {q9, q10}, [r4]
vbit q2, q11, q9
vbit q3, q11, q10
vbit q4, q12, q9
vbit q5, q12, q10
4: // Loop horizontally
vext.8 q7, q2, q3, #4
vext.8 q8, q2, q3, #8
vext.8 q6, q2, q3, #2
vext.8 q9, q2, q3, #10
vadd.i16 q8, q8, q7
vadd.i16 q9, q9, q6
vext.8 q6, q2, q3, #12
vext.8 q7, q2, q3, #6
vadd.i16 q2, q2, q6
vmull.s16 q6, d14, d0[3]
vmlal.s16 q6, d16, d1[0]
vmlal.s16 q6, d18, d1[1]
vmlal.s16 q6, d4, d1[2]
vmull.s16 q7, d15, d0[3]
vmlal.s16 q7, d17, d1[0]
vmlal.s16 q7, d19, d1[1]
vmlal.s16 q7, d5, d1[2]
vext.8 q8, q4, q5, #4
vext.8 q10, q4, q5, #8
vext.8 q9, q4, q5, #2
vext.8 q2, q4, q5, #10
vadd.i16 q10, q10, q8
vadd.i16 q2, q2, q9
vext.8 q8, q4, q5, #12
vext.8 q9, q4, q5, #6
vadd.i16 q4, q4, q8
vmull.s16 q8, d18, d0[3]
vmlal.s16 q8, d20, d1[0]
vmlal.s16 q8, d4, d1[1]
vmlal.s16 q8, d8, d1[2]
vmull.s16 q9, d19, d0[3]
vmlal.s16 q9, d21, d1[0]
vmlal.s16 q9, d5, d1[1]
vmlal.s16 q9, d9, d1[2]
vmvn.i16 q10, #0x8000 // 0x7fff = (1 << 15) - 1
vadd.i32 q6, q6, q14
vadd.i32 q7, q7, q14
vadd.i32 q8, q8, q14
vadd.i32 q9, q9, q14
vrshl.s32 q6, q6, q13
vrshl.s32 q7, q7, q13
vrshl.s32 q8, q8, q13
vrshl.s32 q9, q9, q13
vqmovun.s32 d12, q6
vqmovun.s32 d13, q7
vqmovun.s32 d14, q8
vqmovun.s32 d15, q9
vmin.u16 q6, q6, q10
vmin.u16 q7, q7, q10
vsub.i16 q6, q6, q15
vsub.i16 q7, q7, q15
subs r5, r5, #8
vst1.16 {q6}, [r0, :128]!
vst1.16 {q7}, [r12, :128]!
ble 9f
tst r7, #2 // LR_HAVE_RIGHT
vmov q2, q3
vmov q4, q5
vld1.16 {q3}, [r2]!
vld1.16 {q5}, [lr]!
bne 4b // If we don't need to pad, just keep filtering.
b 3b // If we need to pad, check how many pixels we have left.
9:
subs r6, r6, #2
ble 0f
// Jump to the next row and loop horizontally
add r0, r0, r10
add r12, r12, r10
add r2, r2, r3
add lr, lr, r3
mov r5, r8
b 1b
0:
vpop {q4-q7}
pop {r4-r11,pc}
endfunc
// void dav1d_wiener_filter_v_16bpc_neon(pixel *dst, ptrdiff_t stride,
// const int16_t *mid, int w, int h,
// const int16_t fv[7], enum LrEdgeFlags edges,
// ptrdiff_t mid_stride, const int bitdepth_max);
function wiener_filter_v_16bpc_neon, export=1
push {r4-r7,lr}
vpush {q4-q5}
ldrd r4, r5, [sp, #52]
ldrd r6, r7, [sp, #60]
ldr lr, [sp, #68] // bitdepth_max
vld1.16 {q0}, [r5, :128]
vdup.16 q5, lr
clz lr, lr
sub lr, lr, #11 // round_bits_v
vdup.32 q4, lr
mov lr, r4
vneg.s32 q4, q4 // -round_bits_v
// Calculate the number of rows to move back when looping vertically
mov r12, r4
tst r6, #4 // LR_HAVE_TOP
beq 0f
sub r2, r2, r7, lsl #1
add r12, r12, #2
0:
tst r6, #8 // LR_HAVE_BOTTOM
beq 1f
add r12, r12, #2
1: // Start of horizontal loop; start one vertical filter slice.
// Load rows into q8-q11 and pad properly.
tst r6, #4 // LR_HAVE_TOP
vld1.16 {q8}, [r2, :128], r7
beq 2f
// LR_HAVE_TOP
vld1.16 {q10}, [r2, :128], r7
vmov q9, q8
vld1.16 {q11}, [r2, :128], r7
b 3f
2: // !LR_HAVE_TOP
vmov q9, q8
vmov q10, q8
vmov q11, q8
3:
cmp r4, #4
blt 5f
// Start filtering normally; fill in q12-q14 with unique rows.
vld1.16 {q12}, [r2, :128], r7
vld1.16 {q13}, [r2, :128], r7
vld1.16 {q14}, [r2, :128], r7
4:
.macro filter compare
subs r4, r4, #1
// Interleaving the mul/mla chains actually hurts performance
// significantly on Cortex A53, thus keeping mul/mla tightly
// chained like this.
vmull.s16 q2, d16, d0[0]
vmlal.s16 q2, d18, d0[1]
vmlal.s16 q2, d20, d0[2]
vmlal.s16 q2, d22, d0[3]
vmlal.s16 q2, d24, d1[0]
vmlal.s16 q2, d26, d1[1]
vmlal.s16 q2, d28, d1[2]
vmull.s16 q3, d17, d0[0]
vmlal.s16 q3, d19, d0[1]
vmlal.s16 q3, d21, d0[2]
vmlal.s16 q3, d23, d0[3]
vmlal.s16 q3, d25, d1[0]
vmlal.s16 q3, d27, d1[1]
vmlal.s16 q3, d29, d1[2]
vrshl.s32 q2, q2, q4 // round_bits_v
vrshl.s32 q3, q3, q4
vqmovun.s32 d4, q2
vqmovun.s32 d5, q3
vmin.u16 q2, q2, q5 // bitdepth_max
vst1.16 {q2}, [r0, :128], r1
.if \compare
cmp r4, #4
.else
ble 9f
.endif
vmov q8, q9
vmov q9, q10
vmov q10, q11
vmov q11, q12
vmov q12, q13
vmov q13, q14
.endm
filter 1
blt 7f
vld1.16 {q14}, [r2, :128], r7
b 4b
5: // Less than 4 rows in total; not all of q12-q13 are filled yet.
tst r6, #8 // LR_HAVE_BOTTOM
beq 6f
// LR_HAVE_BOTTOM
cmp r4, #2
// We load at least 2 rows in all cases.
vld1.16 {q12}, [r2, :128], r7
vld1.16 {q13}, [r2, :128], r7
bgt 53f // 3 rows in total
beq 52f // 2 rows in total
51: // 1 row in total, q11 already loaded, load edge into q12-q14.
vmov q13, q12
b 8f
52: // 2 rows in total, q11 already loaded, load q12 with content data
// and 2 rows of edge.
vld1.16 {q14}, [r2, :128], r7
vmov q15, q14
b 8f
53:
// 3 rows in total, q11 already loaded, load q12 and q13 with content
// and 2 rows of edge.
vld1.16 {q14}, [r2, :128], r7
vld1.16 {q15}, [r2, :128], r7
vmov q1, q15
b 8f
6:
// !LR_HAVE_BOTTOM
cmp r4, #2
bgt 63f // 3 rows in total
beq 62f // 2 rows in total
61: // 1 row in total, q11 already loaded, pad that into q12-q14.
vmov q12, q11
vmov q13, q11
vmov q14, q11
b 8f
62: // 2 rows in total, q11 already loaded, load q12 and pad that into q12-q15.
vld1.16 {q12}, [r2, :128], r7
vmov q13, q12
vmov q14, q12
vmov q15, q12
b 8f
63:
// 3 rows in total, q11 already loaded, load q12 and q13 and pad q13 into q14-q15,q1.
vld1.16 {q12}, [r2, :128], r7
vld1.16 {q13}, [r2, :128], r7
vmov q14, q13
vmov q15, q13
vmov q1, q13
b 8f
7:
// All registers up to q13 are filled already, 3 valid rows left.
// < 4 valid rows left; fill in padding and filter the last
// few rows.
tst r6, #8 // LR_HAVE_BOTTOM
beq 71f
// LR_HAVE_BOTTOM; load 2 rows of edge.
vld1.16 {q14}, [r2, :128], r7
vld1.16 {q15}, [r2, :128], r7
vmov q1, q15
b 8f
71:
// !LR_HAVE_BOTTOM, pad 3 rows
vmov q14, q13
vmov q15, q13
vmov q1, q13
8: // At this point, all registers up to q14-q15,q1 are loaded with
// edge/padding (depending on how many rows are left).
filter 0 // This branches to 9f when done
vmov q14, q15
vmov q15, q1
b 8b
9: // End of one vertical slice.
subs r3, r3, #8
ble 0f
// Move pointers back up to the top and loop horizontally.
mls r0, r1, lr, r0
mls r2, r7, r12, r2
add r0, r0, #16
add r2, r2, #16
mov r4, lr
b 1b
0:
vpop {q4-q5}
pop {r4-r7,pc}
.purgem filter
endfunc
#include "looprestoration_tmpl.S"
// void dav1d_sgr_box3_row_h_16bpc_neon(int32_t *sumsq, int16_t *sum,
// const pixel (*left)[4],
// const pixel *src, const int w,
// const enum LrEdgeFlags edges);
function sgr_box3_row_h_16bpc_neon, export=1
push {r4-r5,lr}
ldrd r4, r5, [sp, #12]
add r4, r4, #2 // w += 2
tst r5, #1 // LR_HAVE_LEFT
beq 1f
cmp r2, #0
bne 0f
// LR_HAVE_LEFT && left == NULL
sub r3, r3, #4
vld1.8 {q0, q1}, [r3]!
b 2f
0:
// LR_HAVE_LEFT, left != NULL
vld1.8 {q0, q1}, [r3]!
vld1.16 {d5}, [r2]
// Move r3 back to account for the last 2 pixels we loaded earlier,
// which we'll shift out.
sub r3, r3, #4
vext.8 q1, q0, q1, #12
vext.8 q0, q2, q0, #12
b 2f
1:
vld1.8 {q0, q1}, [r3]!
// !LR_HAVE_LEFT, fill q1 with the leftmost pixel
// and shift q0/q1 to have 2x the first pixel at the front.
vdup.16 q2, d0[0]
// Move r3 back to account for the last 2 pixels we loaded before,
// which we shifted out.
sub r3, r3, #4
vext.8 q1, q0, q1, #12
vext.8 q0, q2, q0, #12
2:
tst r5, #2 // LR_HAVE_RIGHT
bne 4f
// If we'll need to pad the right edge, load that pixel to pad with
// here since we can find it pretty easily from here.
sub lr, r4, #(2 + 16 - 2 + 1)
lsl lr, lr, #1
ldrh lr, [r3, lr]
// Fill q14 with the right padding pixel
vdup.16 q14, lr
3: // !LR_HAVE_RIGHT
// Check whether we need to pad the right edge
cmp r4, #10
bge 4f // If w >= 10, all used input pixels are valid
// 1 <= w < 10, w pixels valid in q0-q1. For w=9, this ends up called
// again; it's not strictly needed in those cases (we pad enough here),
// but keeping the code as simple as possible.
// Insert padding in q0.h[w] onwards
movrel_local lr, right_ext_mask
sub lr, lr, r4, lsl #1
vld1.8 {q12, q13}, [lr]
vbit q0, q14, q12
vbit q1, q14, q13
4: // Loop horizontally
vext.8 q8, q0, q1, #2
vext.8 q9, q0, q1, #4
vadd.i16 q2, q0, q8
vmull.u16 q12, d0, d0
vmlal.u16 q12, d16, d16
vmlal.u16 q12, d18, d18
vadd.i16 q2, q2, q9
vmull.u16 q13, d1, d1
vmlal.u16 q13, d17, d17
vmlal.u16 q13, d19, d19
subs r4, r4, #8
vst1.16 {q2}, [r1, :128]!
vst1.32 {q12, q13}, [r0, :128]!
ble 9f
tst r5, #2 // LR_HAVE_RIGHT
vmov q0, q1
vld1.16 {q1}, [r3]!
bne 4b // If we don't need to pad, just keep summing.
b 3b // If we need to pad, check how many pixels we have left.
9:
pop {r4-r5,pc}
endfunc
// void dav1d_sgr_box5_row_h_16bpc_neon(int32_t *sumsq, int16_t *sum,
// const pixel (*left)[4],
// const pixel *src, const int w,
// const enum LrEdgeFlags edges);
function sgr_box5_row_h_16bpc_neon, export=1
push {r4-r5,lr}
ldrd r4, r5, [sp, #12]
add r4, r4, #2 // w += 2
tst r5, #1 // LR_HAVE_LEFT
beq 1f
cmp r2, #0
bne 0f
// LR_HAVE_LEFT && left == NULL
sub r3, r3, #6
vld1.8 {q0, q1}, [r3]!
b 2f
0:
// LR_HAVE_LEFT, left != NULL
vld1.8 {q0, q1}, [r3]!
vld1.16 {d5}, [r2]
// Move r3 back to account for the last 2 pixels we loaded earlier,
// which we'll shift out.
sub r3, r3, #6
vext.8 q1, q0, q1, #10
vext.8 q0, q2, q0, #10
b 2f
1:
vld1.8 {q0, q1}, [r3]!
// !LR_HAVE_LEFT, fill q1 with the leftmost pixel
// and shift q0/q1 to have 3x the first pixel at the front.
vdup.16 q2, d0[0]
// Move r3 back to account for the last 3 pixels we loaded before,
// which we shifted out.
sub r3, r3, #6
vext.8 q1, q0, q1, #10
vext.8 q0, q2, q0, #10
2:
tst r5, #2 // LR_HAVE_RIGHT
bne 4f
// If we'll need to pad the right edge, load that pixel to pad with
// here since we can find it pretty easily from here.
sub lr, r4, #(2 + 16 - 3 + 1)
lsl lr, lr, #1
ldrh lr, [r3, lr]
// Fill q14 with the right padding pixel
vdup.16 q14, lr
3: // !LR_HAVE_RIGHT
// Check whether we need to pad the right edge
cmp r4, #11
bge 4f // If w >= 11, all used input pixels are valid
// 1 <= w < 11, w+1 pixels valid in q0-q1. For w=9 or w=10,
// this ends up called again; it's not strictly needed in those
// cases (we pad enough here), but keeping the code as simple as possible.
// Insert padding in q0.h[w+1] onwards; fuse the +1 into the
// buffer pointer.
movrel_local lr, right_ext_mask, -2
sub lr, lr, r4, lsl #1
vld1.8 {q12, q13}, [lr]
vbit q0, q14, q12
vbit q1, q14, q13
4: // Loop horizontally
vext.8 q8, q0, q1, #2
vext.8 q9, q0, q1, #4
vadd.i16 q2, q0, q8
vmull.u16 q12, d0, d0
vmlal.u16 q12, d16, d16
vmlal.u16 q12, d18, d18
vadd.i16 q2, q2, q9
vmull.u16 q13, d1, d1
vmlal.u16 q13, d17, d17
vmlal.u16 q13, d19, d19
vext.8 q8, q0, q1, #6
vext.8 q9, q0, q1, #8
vadd.i16 q2, q2, q8
vmlal.u16 q12, d16, d16
vmlal.u16 q12, d1, d1
vadd.i16 q2, q2, q9
vmlal.u16 q13, d17, d17
vmlal.u16 q13, d19, d19
subs r4, r4, #8
vst1.16 {q2}, [r1, :128]!
vst1.32 {q12, q13}, [r0, :128]!
ble 9f
tst r5, #2 // LR_HAVE_RIGHT
vmov q0, q1
vld1.16 {q1}, [r3]!
bne 4b // If we don't need to pad, just keep summing.
b 3b // If we need to pad, check how many pixels we have left.
9:
pop {r4-r5,pc}
endfunc
// void dav1d_sgr_box35_row_h_16bpc_neon(int32_t *sumsq3, int16_t *sum3,
// int32_t *sumsq5, int16_t *sum5,
// const pixel (*left)[4],
// const pixel *src, const int w,
// const enum LrEdgeFlags edges);
function sgr_box35_row_h_16bpc_neon, export=1
push {r4-r7,lr}
ldrd r4, r5, [sp, #20]
ldrd r6, r7, [sp, #28]
add r6, r6, #2 // w += 2
tst r7, #1 // LR_HAVE_LEFT
beq 1f
cmp r4, #0
bne 0f
// LR_HAVE_LEFT && left == NULL
sub r5, r5, #6
vld1.8 {q0, q1}, [r5]!
b 2f
0:
// LR_HAVE_LEFT, left != NULL
vld1.8 {q0, q1}, [r5]!
vld1.16 {d5}, [r4]
// Move r3 back to account for the last 2 pixels we loaded earlier,
// which we'll shift out.
sub r5, r5, #6
vext.8 q1, q0, q1, #10
vext.8 q0, q2, q0, #10
b 2f
1:
vld1.8 {q0, q1}, [r5]!
// !LR_HAVE_LEFT, fill q1 with the leftmost pixel
// and shift q0/q1 to have 3x the first pixel at the front.
vdup.16 q2, d0[0]
// Move r3 back to account for the last 3 pixels we loaded before,
// which we shifted out.
sub r5, r5, #6
vext.8 q1, q0, q1, #10
vext.8 q0, q2, q0, #10
2:
tst r7, #2 // LR_HAVE_RIGHT
bne 4f
// If we'll need to pad the right edge, load that pixel to pad with
// here since we can find it pretty easily from here.
sub lr, r6, #(2 + 16 - 3 + 1)
lsl lr, lr, #1
ldrh lr, [r5, lr]
// Fill q14 with the right padding pixel
vdup.16 q14, lr
3: // !LR_HAVE_RIGHT
// Check whether we need to pad the right edge
cmp r6, #11
bge 4f // If w >= 11, all used input pixels are valid
// 1 <= w < 11, w+1 pixels valid in q0-q1. For w=9 or w=10,
// this ends up called again; it's not strictly needed in those
// cases (we pad enough here), but keeping the code as simple as possible.
// Insert padding in q0.h[w+1] onwards; fuse the +1 into the
// buffer pointer.
movrel_local lr, right_ext_mask, -2
sub lr, lr, r6, lsl #1
vld1.8 {q12, q13}, [lr]
vbit q0, q14, q12
vbit q1, q14, q13
4: // Loop horizontally
vext.8 q8, q0, q1, #2
vext.8 q9, q0, q1, #4
vext.8 q10, q0, q1, #6
vext.8 q11, q0, q1, #8
vadd.i16 q2, q8, q9
vadd.i16 q3, q0, q11
vadd.i16 q2, q2, q10
vmull.u16 q12, d16, d16
vmlal.u16 q12, d18, d18
vmlal.u16 q12, d20, d20
vmull.u16 q13, d17, d17
vmlal.u16 q13, d19, d19
vmlal.u16 q13, d21, d21
vadd.i16 q3, q3, q2
vst1.16 {q2}, [r1, :128]!
vst1.32 {q12, q13}, [r0, :128]!
vmlal.u16 q12, d0, d0
vmlal.u16 q12, d22, d22
vmlal.u16 q13, d1, d1
vmlal.u16 q13, d23, d23
subs r6, r6, #8
vst1.16 {q3}, [r3, :128]!
vst1.32 {q12, q13}, [r2, :128]!
ble 9f
tst r7, #2 // LR_HAVE_RIGHT
vmov q0, q1
vld1.16 {q1}, [r5]!
bne 4b // If we don't need to pad, just keep summing.
b 3b // If we need to pad, check how many pixels we have left.
9:
pop {r4-r7,pc}
endfunc
sgr_funcs 16