ref: 2c09aaa4d7e0f96ca5a4cdefb6a04d0cf7e0bf4b
parent: 7ebcb777b99b954c8570141c469418371b3c2e75
author: Martin Storsjö <[email protected]>
date: Tue Sep 15 14:48:19 EDT 2020
arm32: looprestoration: NEON implementation of wiener filter for 16 bpc
Checkasm benchmarks: Cortex A7 A8 A53 A72 A73
wiener_chroma_10bpc_neon: 385312.5 165772.7 184308.2 122311.2 126050.2
wiener_chroma_12bpc_neon: 385296.7 165538.0 184438.2 122290.5 126205.3
wiener_luma_10bpc_neon: 385318.5 165985.3 184147.4 122311.1 126168.4
wiener_luma_12bpc_neon: 385316.3 165819.1 184484.7 122304.4 125982.4
The corresponding numbers for arm64 for comparison:
Cortex A53 A72 A73
wiener_chroma_10bpc_neon: 176319.7 125992.1 128162.4
wiener_chroma_12bpc_neon: 176386.2 125986.4 128343.8
wiener_luma_10bpc_neon: 176174.0 126001.7 128227.8
wiener_luma_12bpc_neon: 176176.5 125992.1 128204.8
The arm32 version actually seems to run marginally faster than the arm64
one on A72 and A73. I believe this is because the arm64 code is tuned
for A53 (which makes it a bit slower on other cores), but the arm32 code
can't be tuned exactly the same way due to fewer registers being available.
--- /dev/null
+++ b/src/arm/32/looprestoration16.S
@@ -1,0 +1,720 @@
+/*
+ * 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"
+
+// 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]+ 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
+
+ // Clear the last unused element of q0, to allow filtering a single
+ // pixel with one plain vmul+vpadd.
+ mov r12, #0
+ vmov.16 d1[3], r12
+
+ // 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 width from mid_stride
+ sub r10, r10, r5, lsl #1
+
+ // For w >= 8, we read (w+5)&~7+8 pixels, for w < 8 we read 16 pixels.
+ cmp r5, #8
+ add r11, r5, #13
+ bic r11, r11, #7
+ bge 1f
+ mov r11, #16
+1:
+ 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 byte 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
+ // If we'll have to pad the right edge we need to quit early here.
+ cmp r5, #11
+ bge 4f // If w >= 11, all used input pixels are valid
+ cmp r5, #7
+ bge 5f // If w >= 7, we can filter 4 pixels
+ b 6f
+
+4: // Loop horizontally
+ vext.8 q10, q2, q3, #6
+ vext.8 q8, q2, q3, #2
+ vext.8 q9, q2, q3, #4
+ vshll.u16 q6, d20, #7
+ vshll.u16 q7, d21, #7
+ vmlal.s16 q6, d4, d0[0]
+ vmlal.s16 q6, d16, d0[1]
+ vmlal.s16 q6, d18, d0[2]
+ vmlal.s16 q6, d20, d0[3]
+ vmlal.s16 q7, d5, d0[0]
+ vmlal.s16 q7, d17, d0[1]
+ vmlal.s16 q7, d19, d0[2]
+ vmlal.s16 q7, d21, d0[3]
+ vext.8 q8, q2, q3, #8
+ vext.8 q9, q2, q3, #10
+ vext.8 q10, q2, q3, #12
+ vmlal.s16 q6, d16, d1[0]
+ vmlal.s16 q6, d18, d1[1]
+ vmlal.s16 q6, d20, d1[2]
+ vmlal.s16 q7, d17, d1[0]
+ vmlal.s16 q7, d19, d1[1]
+ vmlal.s16 q7, d21, d1[2]
+ vext.8 q10, q4, q5, #6
+ vext.8 q2, q4, q5, #2
+ vshll.u16 q8, d20, #7
+ vshll.u16 q9, d21, #7
+ vmlal.s16 q8, d8, d0[0]
+ vmlal.s16 q8, d4, d0[1]
+ vmlal.s16 q8, d20, d0[3]
+ vmlal.s16 q9, d9, d0[0]
+ vmlal.s16 q9, d5, d0[1]
+ vmlal.s16 q9, d21, d0[3]
+ vext.8 q2, q4, q5, #4
+ vext.8 q10, q4, q5, #8
+ vmlal.s16 q8, d4, d0[2]
+ vmlal.s16 q8, d20, d1[0]
+ vmlal.s16 q9, d5, d0[2]
+ vmlal.s16 q9, d21, d1[0]
+ vext.8 q2, q4, q5, #10
+ vext.8 q10, q4, q5, #12
+ vmlal.s16 q8, d4, d1[1]
+ vmlal.s16 q8, d20, d1[2]
+ vmlal.s16 q9, d5, d1[1]
+ vmlal.s16 q9, d21, 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
+ vst1.16 {q6}, [r0, :128]!+ vst1.16 {q7}, [r12, :128]!+
+ subs r5, r5, #8
+ 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.
+
+5: // Filter 4 pixels, 7 <= w < 11
+.macro filter_4
+ vext.8 d18, d4, d5, #6
+ vext.8 d16, d4, d5, #2
+ vext.8 d17, d4, d5, #4
+ vext.8 d19, d5, d6, #2
+ vext.8 d20, d5, d6, #4
+ vshll.u16 q6, d18, #7
+ vmlal.s16 q6, d4, d0[0]
+ vmlal.s16 q6, d16, d0[1]
+ vmlal.s16 q6, d17, d0[2]
+ vmlal.s16 q6, d18, d0[3]
+ vmlal.s16 q6, d5, d1[0]
+ vmlal.s16 q6, d19, d1[1]
+ vmlal.s16 q6, d20, d1[2]
+
+ vext.8 d18, d8, d9, #6
+ vext.8 d16, d8, d9, #2
+ vext.8 d17, d8, d9, #4
+ vext.8 d19, d9, d10, #2
+ vext.8 d20, d9, d10, #4
+ vshll.u16 q7, d18, #7
+ vmlal.s16 q7, d8, d0[0]
+ vmlal.s16 q7, d16, d0[1]
+ vmlal.s16 q7, d17, d0[2]
+ vmlal.s16 q7, d18, d0[3]
+ vmlal.s16 q7, d9, d1[0]
+ vmlal.s16 q7, d19, d1[1]
+ vmlal.s16 q7, d20, d1[2]
+
+ vmvn.i16 q10, #0x8000 // 0x7fff = (1 << 15) - 1
+ vadd.i32 q6, q6, q14
+ vadd.i32 q7, q7, q14
+ vrshl.s32 q6, q6, q13
+ vrshl.s32 q7, q7, q13
+ vqmovun.s32 d12, q6
+ vqmovun.s32 d13, q7
+ vmin.u16 q6, q6, q10
+ vsub.i16 q6, q6, q15
+.endm
+ filter_4
+ vst1.16 {d12}, [r0, :64]!+ vst1.16 {d13}, [r12, :64]!+
+ subs r5, r5, #4 // 3 <= w < 7
+ vext.8 q2, q2, q3, #8
+ vext.8 q3, q3, q3, #8
+ vext.8 q4, q4, q5, #8
+ vext.8 q5, q5, q5, #8
+
+6: // Pad the right edge and filter the last few pixels.
+ // w < 7, w+3 pixels valid in q2-q3
+ cmp r5, #5
+ blt 7f
+ bgt 8f
+ // w == 5, 8 pixels valid in q2, q3 invalid
+ vmov q3, q11
+ vmov q5, q12
+ b 88f
+
+7: // 1 <= w < 5, 4-7 pixels valid in q2
+ sub r9, r5, #1
+ // r9 = (pixels valid - 4)
+ adr r11, L(variable_shift_tbl)
+ ldr r9, [r11, r9, lsl #2]
+ add r11, r11, r9
+ vmov q3, q11
+ vmov q5, q12
+ bx r11
+
+ .align 2
+L(variable_shift_tbl):
+ .word 44f - L(variable_shift_tbl) + CONFIG_THUMB
+ .word 55f - L(variable_shift_tbl) + CONFIG_THUMB
+ .word 66f - L(variable_shift_tbl) + CONFIG_THUMB
+ .word 77f - L(variable_shift_tbl) + CONFIG_THUMB
+
+44: // 4 pixels valid in q2/q4, fill the high half with padding.
+ vmov d5, d6
+ vmov d9, d10
+ b 88f
+ // Shift q2 right, shifting out invalid pixels,
+ // shift q2 left to the original offset, shifting in padding pixels.
+55: // 5 pixels valid
+ vext.8 q2, q2, q2, #10
+ vext.8 q2, q2, q3, #6
+ vext.8 q4, q4, q4, #10
+ vext.8 q4, q4, q5, #6
+ b 88f
+66: // 6 pixels valid
+ vext.8 q2, q2, q2, #12
+ vext.8 q2, q2, q3, #4
+ vext.8 q4, q4, q4, #12
+ vext.8 q4, q4, q5, #4
+ b 88f
+77: // 7 pixels valid
+ vext.8 q2, q2, q2, #14
+ vext.8 q2, q2, q3, #2
+ vext.8 q4, q4, q4, #14
+ vext.8 q4, q4, q5, #2
+ b 88f
+
+8: // w > 5, w == 6, 9 pixels valid in q2-q3, 1 pixel valid in q3
+ vext.8 q3, q3, q3, #2
+ vext.8 q3, q3, q11, #14
+ vext.8 q5, q5, q5, #2
+ vext.8 q5, q5, q12, #14
+
+88:
+ // w < 7, q2-q3 padded properly
+ cmp r5, #4
+ blt 888f
+
+ // w >= 4, filter 4 pixels
+ filter_4
+ vst1.16 {d12}, [r0, :64]!+ vst1.16 {d13}, [r12, :64]!+ subs r5, r5, #4 // 0 <= w < 4
+ vext.8 q2, q2, q3, #8
+ vext.8 q4, q4, q5, #8
+ beq 9f
+888: // 1 <= w < 4, filter 1 pixel at a time
+ vmull.s16 q6, d4, d0
+ vmull.s16 q7, d5, d1
+ vmull.s16 q8, d8, d0
+ vmull.s16 q9, d9, d1
+ vadd.i32 q6, q7
+ vadd.i32 q8, q9
+ vpadd.i32 d12, d12, d13
+ vpadd.i32 d13, d16, d17
+ vdup.16 d14, d4[3]
+ vdup.16 d15, d8[3]
+ vpadd.i32 d12, d12, d13
+ vtrn.16 d14, d15
+ vadd.i32 d12, d12, d28
+ vshll.u16 q7, d14, #7
+ vmvn.i16 d20, #0x8000 // 0x7fff = (1 << 15) - 1
+ vadd.i32 d12, d12, d14
+ vrshl.s32 d12, d12, d26
+ vqmovun.s32 d12, q6
+ vmin.u16 d12, d12, d20
+ vsub.i16 d12, d12, d30
+ vst1.16 {d12[0]}, [r0, :16]!+ vst1.16 {d12[1]}, [r12, :16]!+ subs r5, r5, #1
+ vext.8 q2, q2, q3, #2
+ vext.8 q4, q4, q5, #2
+ bgt 888b
+
+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}+.purgem filter_4
+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
+ vmov.i16 q1, #0
+ mov r12, #128
+ vld1.16 {q0}, [r5]+ vdup.16 q5, lr
+ clz lr, lr
+ vmov.i16 d2[3], r12
+ sub lr, lr, #11 // round_bits_v
+ vadd.i16 q0, q0, q1
+ 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], 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
+
+// void dav1d_copy_narrow_16bpc_neon(pixel *dst, ptrdiff_t stride,
+// const pixel *src, int w, int h);
+function copy_narrow_16bpc_neon, export=1
+ push {r4,lr}+ ldr r4, [sp, #8]
+ adr r12, L(copy_narrow_tbl)
+ ldr r3, [r12, r3, lsl #2]
+ add r12, r12, r3
+ bx r12
+
+ .align 2
+L(copy_narrow_tbl):
+ .word 0
+ .word 10f - L(copy_narrow_tbl) + CONFIG_THUMB
+ .word 20f - L(copy_narrow_tbl) + CONFIG_THUMB
+ .word 30f - L(copy_narrow_tbl) + CONFIG_THUMB
+ .word 40f - L(copy_narrow_tbl) + CONFIG_THUMB
+ .word 50f - L(copy_narrow_tbl) + CONFIG_THUMB
+ .word 60f - L(copy_narrow_tbl) + CONFIG_THUMB
+ .word 70f - L(copy_narrow_tbl) + CONFIG_THUMB
+
+10:
+ add r3, r0, r1
+ lsl r1, r1, #1
+18:
+ subs r4, r4, #8
+ blt 110f
+ vld1.16 {q0}, [r2, :128]!+ vst1.16 {d0[0]}, [r0, :16], r1+ vst1.16 {d0[1]}, [r3, :16], r1+ vst1.16 {d0[2]}, [r0, :16], r1+ vst1.16 {d0[3]}, [r3, :16], r1+ vst1.16 {d1[0]}, [r0, :16], r1+ vst1.16 {d1[1]}, [r3, :16], r1+ vst1.16 {d1[2]}, [r0, :16], r1+ vst1.16 {d1[3]}, [r3, :16], r1+ ble 0f
+ b 18b
+110:
+ add r4, r4, #8
+ asr r1, r1, #1
+11:
+ subs r4, r4, #1
+ vld1.16 {d0[]}, [r2]!+ vst1.16 {d0[0]}, [r0], r1+ bgt 11b
+0:
+ pop {r4,pc}+
+20:
+ add r3, r0, r1
+ lsl r1, r1, #1
+24:
+ subs r4, r4, #4
+ blt 210f
+ vld1.32 {q0}, [r2, :128]!+ vst1.32 {d0[0]}, [r0, :32], r1+ vst1.32 {d0[1]}, [r3, :32], r1+ vst1.32 {d1[0]}, [r0, :32], r1+ vst1.32 {d1[1]}, [r3, :32], r1+ ble 0f
+ b 24b
+210:
+ add r4, r4, #4
+ asr r1, r1, #1
+22:
+ subs r4, r4, #1
+ vld1.32 {d0[]}, [r2, :32]!+ vst1.32 {d0[0]}, [r0, :32], r1+ bgt 22b
+0:
+ pop {r4,pc}+
+30:
+ ldr r3, [r2]
+ ldrh r12, [r2, #4]
+ add r2, r2, #6
+ subs r4, r4, #1
+ str r3, [r0]
+ strh r12, [r0, #4]
+ add r0, r0, r1
+ bgt 30b
+ pop {r4,pc}+
+40:
+ add r3, r0, r1
+ lsl r1, r1, #1
+42:
+ subs r4, r4, #2
+ blt 41f
+ vld1.16 {q0}, [r2, :128]!+ vst1.16 {d0}, [r0, :64], r1+ vst1.16 {d1}, [r3, :64], r1+ ble 0f
+ b 42b
+41:
+ vld1.16 {d0}, [r2, :64]+ vst1.16 {d0}, [r0, :64]+0:
+ pop {r4,pc}+
+50:
+ vld1.16 {d0}, [r2]+ ldrh r12, [r2, #8]
+ add r2, r2, #10
+ subs r4, r4, #1
+ vst1.16 {d0}, [r0]+ strh r12, [r0, #8]
+ add r0, r0, r1
+ bgt 50b
+ pop {r4,pc}+
+60:
+ vld1.16 {d0}, [r2]+ ldr r12, [r2, #8]
+ add r2, r2, #12
+ subs r4, r4, #1
+ vst1.16 {d0}, [r0]+ str r12, [r0, #8]
+ add r0, r0, r1
+ bgt 60b
+ pop {r4,pc}+
+70:
+ vld1.16 {d0}, [r2]+ ldr r12, [r2, #8]
+ ldrh lr, [r2, #12]
+ add r2, r2, #14
+ subs r4, r4, #1
+ vst1.16 {d0}, [r0]+ str r12, [r0, #8]
+ strh lr, [r0, #12]
+ add r0, r0, r1
+ bgt 70b
+ pop {r4,pc}+endfunc
--- a/src/arm/looprestoration_init_tmpl.c
+++ b/src/arm/looprestoration_init_tmpl.c
@@ -29,7 +29,6 @@
#include "src/looprestoration.h"
#include "src/tables.h"
-#if BITDEPTH == 8 || ARCH_AARCH64
// The 8bpc version calculates things slightly differently than the reference
// C version. That version calculates roughly this:
// int16_t sum = 0;
@@ -105,6 +104,7 @@
}
}
+#if BITDEPTH == 8 || ARCH_AARCH64
void BF(dav1d_sgr_box3_h, neon)(int32_t *sumsq, int16_t *sum,
const pixel (*left)[4],
const pixel *src, const ptrdiff_t stride,
@@ -290,8 +290,8 @@
if (!(flags & DAV1D_ARM_CPU_FLAG_NEON)) return;
-#if BITDEPTH == 8 || ARCH_AARCH64
c->wiener = wiener_filter_neon;
+#if BITDEPTH == 8 || ARCH_AARCH64
if (bpc <= 10)
c->selfguided = sgr_filter_neon;
#endif
--- a/src/meson.build
+++ b/src/meson.build
@@ -147,6 +147,7 @@
if dav1d_bitdepths.contains('16')libdav1d_sources_asm += files(
+ 'arm/32/looprestoration16.S',
'arm/32/mc16.S',
)
endif