[FFmpeg-devel] [PATCH 2/5] swscale/aarch64: Add rgb24 to yuv implementation
Martin Storsjö
martin at martin.st
Wed Jun 5 11:16:24 EEST 2024
On Tue, 4 Jun 2024, Zhao Zhili wrote:
> From: Zhao Zhili <zhilizhao at tencent.com>
>
> Test on Apple M1:
>
> rgb24_to_uv_1080_c: 7.2
> rgb24_to_uv_1080_neon: 5.5
> rgb24_to_uv_1280_c: 8.2
> rgb24_to_uv_1280_neon: 6.2
> rgb24_to_uv_1920_c: 12.5
> rgb24_to_uv_1920_neon: 9.5
>
> rgb24_to_uv_half_540_c: 6.5
> rgb24_to_uv_half_540_neon: 3.0
> rgb24_to_uv_half_640_c: 7.5
> rgb24_to_uv_half_640_neon: 3.2
> rgb24_to_uv_half_960_c: 12.5
> rgb24_to_uv_half_960_neon: 6.0
>
> rgb24_to_y_1080_c: 4.5
> rgb24_to_y_1080_neon: 3.5
> rgb24_to_y_1280_c: 5.2
> rgb24_to_y_1280_neon: 4.2
> rgb24_to_y_1920_c: 8.0
> rgb24_to_y_1920_neon: 6.0
>
> Signed-off-by: Zhao Zhili <zhilizhao at tencent.com>
> ---
> libswscale/aarch64/Makefile | 1 +
> libswscale/aarch64/input.S | 229 +++++++++++++++++++++++++++++++++++
> libswscale/aarch64/swscale.c | 25 ++++
> 3 files changed, 255 insertions(+)
> create mode 100644 libswscale/aarch64/input.S
>
> diff --git a/libswscale/aarch64/Makefile b/libswscale/aarch64/Makefile
> index da1d909561..adfd90a1b6 100644
> --- a/libswscale/aarch64/Makefile
> +++ b/libswscale/aarch64/Makefile
> @@ -3,6 +3,7 @@ OBJS += aarch64/rgb2rgb.o \
> aarch64/swscale_unscaled.o \
>
> NEON-OBJS += aarch64/hscale.o \
> + aarch64/input.o \
> aarch64/output.o \
> aarch64/rgb2rgb_neon.o \
> aarch64/yuv2rgb_neon.o \
> diff --git a/libswscale/aarch64/input.S b/libswscale/aarch64/input.S
> new file mode 100644
> index 0000000000..ee0d223c6e
> --- /dev/null
> +++ b/libswscale/aarch64/input.S
> @@ -0,0 +1,229 @@
> +/*
> + * Copyright (c) 2024 Zhao Zhili <quinkblack at foxmail.com>
> + *
> + * This file is part of FFmpeg.
> + *
> + * FFmpeg is free software; you can redistribute it and/or
> + * modify it under the terms of the GNU Lesser General Public
> + * License as published by the Free Software Foundation; either
> + * version 2.1 of the License, or (at your option) any later version.
> + *
> + * FFmpeg is distributed in the hope that it will be useful,
> + * but WITHOUT ANY WARRANTY; without even the implied warranty of
> + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
> + * Lesser General Public License for more details.
> + *
> + * You should have received a copy of the GNU Lesser General Public
> + * License along with FFmpeg; if not, write to the Free Software
> + * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
> + */
> +
> +#include "libavutil/aarch64/asm.S"
> +
> +.macro rgb24_to_yuv_load_rgb, src
> + ld3 { v16.16b, v17.16b, v18.16b }, [\src]
> + ushll v19.8h, v16.8b, #0 // v19: r
> + ushll v20.8h, v17.8b, #0 // v20: g
> + ushll v21.8h, v18.8b, #0 // v21: b
> + ushll2 v22.8h, v16.16b, #0 // v22: r
> + ushll2 v23.8h, v17.16b, #0 // v23: g
> + ushll2 v24.8h, v18.16b, #0 // v24: b
Doing "ushll #0" is perhaps a bit unusual, the common thing would be
"uxtl" instead. It doesn't matter in practice though, it assembles to the
same instruction anyway.
> +.endm
> +
> +.macro rgb24_to_yuv_product, r, g, b, dst1, dst2, dst, coef0, coef1, coef2, right_shift
> + mov \dst1\().16b, v6.16b // dst1 = const_offset
> + mov \dst2\().16b, v6.16b // dst2 = const_offset
> + smlal \dst1\().4s, \coef0\().4h, \r\().4h // dst1 += rx * r
> + smlal2 \dst2\().4s, \coef0\().8h, \r\().8h // dst2 += rx * r
> + smlal \dst1\().4s, \coef1\().4h, \g\().4h // dst1 += gx * g
> + smlal2 \dst2\().4s, \coef1\().8h, \g\().8h // dst2 += gx * g
> + smlal \dst1\().4s, \coef2\().4h, \b\().4h // dst1 += bx * b
> + smlal2 \dst2\().4s, \coef2\().8h, \b\().8h // dst2 += bx * b
For sequences like this, the Cortex A53 (and iirc at least the A55 too)
has got a fastpath; if you do multiple consequent smlal/smlsl (or regular
mla/mls) into the same register, you actually save a lot of time. E.g.
instead of this:
smlal dst1
smlal dst2
smlal dst1
smlal dst2
smlal dst1
smlal dst2
Do this:
smlal dst1
smlal dst1
smlal dst1
smlal dst2
smlal dst2
smlal dst2
For in-order cores (with this special fastpath - it is indeed a bit
non-obvious) this makes a huge difference, and for out of order cores,
they can reorder it as they prefer anyway (as this is not a very long
instruction sequence).
This makes a massive difference for the in-order cores.
Before: Cortex A53 A72 A73
rgb24_to_y_1920_neon: 6032.7 3385.7 2514.0
After:
rgb24_to_y_1920_neon: 5072.7 3388.2 2522.0
A 19% speedup on A53 with just with this one change, and it makes almost
no difference for the other cores (mostly within measurement noise).
> +function ff_rgb24ToY_neon, export=1
> + cmp w4, #0 // check width > 0
> + b.le 4f
> +
> + ldp w10, w11, [x5], #8 // w10: ry, w11: gy
> + dup v0.8h, w10
> + dup v1.8h, w11
> + ldr w12, [x5] // w12: by
> + dup v2.8h, w12
> +
> + mov w9, #256 // w9 = 1 << (RGB2YUV_SHIFT - 7)
> + movk w9, #8, lsl #16 // w9 += 32 << (RGB2YUV_SHIFT - 1)
> + dup v6.4s, w9 // w9: const_offset
> +
> + mov x2, #0 // w2: i
> + and w3, w4, #0xFFFFFFF0 // w3 = width / 16 * 16
> + cbz w3, 3f
This blindly assumes that if width > 0, then width is also >= 16 at least,
as we always run the SIMD codepath once here. Is this a requirement in
swscale, or should we check whether width >= 16 here too?
> +1:
> + rgb24_to_yuv_load_rgb x1
> + rgb24_to_yuv_product v19, v20, v21, v25, v26, v16, v0, v1, v2, #9
> + rgb24_to_yuv_product v22, v23, v24, v27, v28, v17, v0, v1, v2, #9
> + stp q16, q17, [x0], #32 // store to dst
> +
> + add w2, w2, #16 // i += 16
> + add x1, x1, #48 // src += 48
> + cmp w2, w3 // i < (width / 16 * 16)
> + b.lt 1b
For the in-order cores, we can help with scheduling here too. You can do
the add/add/cmp before the stp. This adds a bit more distance between the
calculation of q17 and the store of it, and adds more distance between the
cmp and the b.lt that depends on it.
Secondly, it's a bit uncommon in SIMD like this, to count upwards
(incrementing i); instead, the common pattern is to keep the width in one
register and count it down.
Here, you can then do "cmp w2, #16", so you don't need to keep a register
with the aligned end value. And for the scalar codepath, when approaching
zero, you can do "subs w2, #1", "b.gt 2b", so you don't need two
instructions for add+cmp.
> + b 3f
> +2:
> + ldrb w13, [x1] // w13: r
> + ldrb w14, [x1, #1] // w14: g
> + ldrb w15, [x1, #2] // w15: b
> +
> + smaddl x13, w13, w10, x9 // x13 = ry * r + const_offset
> + smaddl x13, w14, w11, x13 // x13 += gy * g
> + smaddl x13, w15, w12, x13 // x13 += by * b
> + asr w13, w13, #9 // x13 >>= 9
> + strh w13, [x0], #2 // store to dst
> +
> + add w2, w2, #1 // i++
> + add x1, x1, #3 // src += 3
> +3:
> + cmp w2, w4 // i < width
> + b.lt 2b
> +4:
> + ret
> +endfunc
> +
> +.macro rgb24_load_uv_coeff half
> + add x6, x6, #12
> +
> + ldp w10, w11, [x6], #8 // w10: ru, w11: gu
> + dup v0.8h, w10
> + dup v1.8h, w11
> +
> + ldp w12, w13, [x6], #8 // w12: bu, w13: rv
> + dup v2.8h, w12
> + dup v3.8h, w13
> +
> + ldp w14, w15, [x6], #8 // w14: gv, w15: bv
> + dup v4.8h, w14
> + dup v5.8h, w15
As Remi mentioned, scheduling instructions like these can help a bit. But
here, each "ldp" depends on the updated x6 from the previous one, so it
maybe doesn't make much difference.
> +
> + .if \half
> + mov w9, #512
> + movk w9, #128, lsl #16 // w9: const_offset
> + .else
> + mov w9, #256
> + movk w9, #64, lsl #16 // w9: const_offset
> + .endif
> + dup v6.4s, w9
> +.endm
> +
> +function ff_rgb24ToUV_half_neon, export=1
> + cmp w5, #0 // check width > 0
> + b.le 4f
> +
> + rgb24_load_uv_coeff half=1
> +
> + mov x9, #0 // x9: i
> + and w7, w5, #0xFFFFFFF8 // w7 = width / 8 * 8
> + cbz w7, 3f
> +1:
> + ld3 { v16.16b, v17.16b, v18.16b }, [x3]
> + uaddlp v19.8h, v16.16b // v19: r
> + uaddlp v20.8h, v17.16b // v20: g
> + uaddlp v21.8h, v18.16b // v21: b
> +
> + rgb24_to_yuv_product v19, v20, v21, v22, v23, v16, v0, v1, v2, #10
> + str q16, [x0], #16 // store dst_u
> + rgb24_to_yuv_product v19, v20, v21, v24, v25, v17, v3, v4, v5, #10
> + str q17, [x1], #16 // store dst_v
> +
> + add w9, w9, #8 // i += 8
> + add x3, x3, #48 // src += 48
> + cmp w9, w7 // i < (width * 8 / 8)
> + b.lt 1b
Here, you can also move the both str out to after the add/cmp, like above.
> + b 3f
> +2:
> + ldrb w2, [x3] // w2: r1
> + ldrb w4, [x3, #3] // w4: r2
> + add w2, w2, w4 // w2 = r1 + r2
> +
> + ldrb w4, [x3, #1] // w4: g1
> + ldrb w7, [x3, #4] // w7: g2
> + add w4, w4, w7 // w4 = g1 + g2
> +
> + ldrb w7, [x3, #2] // w7: b1
> + ldrb w8, [x3, #5] // w8: b2
> + add w7, w7, w8 // w7 = b1 + b2
> +
> + umov w8, v6.s[0] // dst_u = const_offset
> + smaddl x8, w2, w10, x8 // dst_u += ru * r
> + smaddl x8, w4, w11, x8 // dst_u += gu * g
> + smaddl x8, w7, w12, x8 // dst_u += bu * b
> + asr x8, x8, #10 // dst_u >>= 10
> + strh w8, [x0], #2 // store dst_u
> +
> + umov w8, v6.s[0] // dst_v = const_offset
> + smaddl x8, w2, w13, x8 // dst_v += rv * r
> + smaddl x8, w4, w14, x8 // dst_v += gv * g
> + smaddl x8, w7, w15, x8 // dst_v += bv * b
> + asr x8, x8, #10 // dst_v >>= 10
> + strh w8, [x1], #2 // store dst_v
> +
> + add w9, w9, #1 // i++
> + add x3, x3, #6 // src += 6
> +3:
> + cmp w9, w5
> + b.lt 2b
> +4:
> + ret
> +endfunc
> +
> +function ff_rgb24ToUV_neon, export=1
> + cmp w5, #0 // check width > 0
> + b.le 4f
> +
> + rgb24_load_uv_coeff half=0
> +
> + mov x2, #0 // w2: i
> + and w4, w5, #0xFFFFFFF0 // w4: width / 16 * 16
> + cbz w4, 3f
> +1:
> + rgb24_to_yuv_load_rgb x3
> + rgb24_to_yuv_product v19, v20, v21, v25, v26, v16, v0, v1, v2, #9
> + rgb24_to_yuv_product v22, v23, v24, v27, v28, v17, v0, v1, v2, #9
> + stp q16, q17, [x0], #32 // store to dst_u
> + rgb24_to_yuv_product v19, v20, v21, v25, v26, v16, v3, v4, v5, #9
> + rgb24_to_yuv_product v22, v23, v24, v27, v28, v17, v3, v4, v5, #9
> + stp q16, q17, [x1], #32 // store to dst_v
If you'd make the second pair of rgb24_to_yuv_product write into v18/v19
instead of v16/v17, you can move them out to after the add/add/cmp. (It
doesn't make much of a measurable difference, but is good scheduling in
general.)
> +
> + add w2, w2, #16 // i += 16
> + add x3, x3, #48 // src += 48
> + cmp w2, w4 // i < (width / 16 * 16)
> + b.lt 1b
> + b 3f
> +2:
> + ldrb w16, [x3] // w16: r
> + ldrb w17, [x3, #1] // w17: g
> + ldrb w4, [x3, #2] // w4: b
> +
> + umov w7, v6.s[0] // w7 = const_offset
SIMD->GPR moves are generally expensive, we shouldn't be doing this
within the loop, if the value is constant. Instead, you should do this in
a prologue to the scalar codepath.
But in this case, the value should already be present in w9, if I
understand the code correctly, so we don't need to fetch it from v6 - and
this is already what you do in ff_rgb24ToY_neon.
// Martin
More information about the ffmpeg-devel
mailing list