[FFmpeg-devel] [PATCH 2/2] swscale/aarch64: add vscale specializations
Martin Storsjö
martin at martin.st
Wed Apr 20 00:12:26 EEST 2022
On Fri, 15 Apr 2022, Swinney, Jonathan wrote:
> This commit adds new code paths for vscale when filterSize is 2, 4, or 8. By
> using specialized code with unrolling to match the filterSize we can improve
> performance.
>
> | (seconds) | c6g | | |
> | ------------| ----- | ----- | ----- |
> | filterSize | 2 | 4 | 8 |
> | original | 0.581 | 0.974 | 1.744 |
> | optimized | 0.399 | 0.569 | 1.052 |
> | improvement | 31.1% | 41.6% | 39.7% |
>
> Signed-off-by: Jonathan Swinney <jswinney at amazon.com>
> ---
> libswscale/aarch64/output.S | 147 +++++++++++++++++++++++++++++++++--
> libswscale/aarch64/swscale.c | 12 +++
> 2 files changed, 153 insertions(+), 6 deletions(-)
>
> diff --git a/libswscale/aarch64/output.S b/libswscale/aarch64/output.S
> index af71de6050..9c99c3bea9 100644
> --- a/libswscale/aarch64/output.S
> +++ b/libswscale/aarch64/output.S
> @@ -21,12 +21,27 @@
> #include "libavutil/aarch64/asm.S"
>
> function ff_yuv2planeX_8_neon, export=1
> +// x0 - const int16_t *filter,
> +// x1 - int filterSize,
> +// x2 - const int16_t **src,
> +// x3 - uint8_t *dest,
> +// x4 - int dstW,
> +// x5 - const uint8_t *dither,
> +// x6 - int offset
> +
> ld1 {v0.8B}, [x5] // load 8x8-bit dither
> cbz w6, 1f // check if offsetting present
> ext v0.8B, v0.8B, v0.8B, #3 // honor offsetting which can be 0 or 3 only
> 1: uxtl v0.8H, v0.8B // extend dither to 16-bit
> ushll v1.4S, v0.4H, #12 // extend dither to 32-bit with left shift by 12 (part 1)
> ushll2 v2.4S, v0.8H, #12 // extend dither to 32-bit with left shift by 12 (part 2)
> + cmp w1, #8 // if filterSize == 8, branch to specialized version
> + b.eq 5f
> + cmp w1, #4 // if filterSize == 4, branch to specialized version
> + b.eq 7f
> + cmp w1, #2 // if filterSize == 2, branch to specialized version
> + b.eq 9f
> +
> mov x7, #0 // i = 0
> 2: mov v3.16B, v1.16B // initialize accumulator part 1 with dithering value
> mov v4.16B, v2.16B // initialize accumulator part 2 with dithering value
> @@ -34,16 +49,15 @@ function ff_yuv2planeX_8_neon, export=1
> mov x9, x2 // srcp = src
> mov x10, x0 // filterp = filter
> 3: ldp x11, x12, [x9], #16 // get 2 pointers: src[j] and src[j+1]
> + ld2r {v16.8H, v17.8H}, [x10], #4 // read 2x16-bit coeff X and Y at filter[j] and filter[j+1]
> add x11, x11, x7, lsl #1 // &src[j ][i]
> add x12, x12, x7, lsl #1 // &src[j+1][i]
> ld1 {v5.8H}, [x11] // read 8x16-bit @ src[j ][i + {0..7}]: A,B,C,D,E,F,G,H
> ld1 {v6.8H}, [x12] // read 8x16-bit @ src[j+1][i + {0..7}]: I,J,K,L,M,N,O,P
> - ld1r {v7.8H}, [x10], #2 // read 1x16-bit coeff X at filter[j ] and duplicate across lanes
> - ld1r {v16.8H}, [x10], #2 // read 1x16-bit coeff Y at filter[j+1] and duplicate across lanes
> - smlal v3.4S, v5.4H, v7.4H // val0 += {A,B,C,D} * X
> - smlal2 v4.4S, v5.8H, v7.8H // val1 += {E,F,G,H} * X
> - smlal v3.4S, v6.4H, v16.4H // val0 += {I,J,K,L} * Y
> - smlal2 v4.4S, v6.8H, v16.8H // val1 += {M,N,O,P} * Y
> + smlal v3.4S, v5.4H, v16.4H // val0 += {A,B,C,D} * X
> + smlal2 v4.4S, v5.8H, v16.8H // val1 += {E,F,G,H} * X
> + smlal v3.4S, v6.4H, v17.4H // val0 += {I,J,K,L} * Y
> + smlal2 v4.4S, v6.8H, v17.8H // val1 += {M,N,O,P} * Y
> subs w8, w8, #2 // tmpfilterSize -= 2
> b.gt 3b // loop until filterSize consumed
>
> @@ -55,4 +69,125 @@ function ff_yuv2planeX_8_neon, export=1
> add x7, x7, #8 // i += 8
> b.gt 2b // loop until width consumed
> ret
> +
> +5: // fs=8
> + ldp x5, x6, [x2] // load 2 pointers: src[j ] and src[j+1]
> + ldp x7, x9, [x2, #16] // load 2 pointers: src[j+2] and src[j+3]
> + ldp x10, x11, [x2, #32] // load 2 pointers: src[j+4] and src[j+5]
> + ldp x12, x13, [x2, #48] // load 2 pointers: src[j+6] and src[j+7]
> +
> + // load 8x16-bit values for filter[j], where j=0..7 and replicated across lanes
> + ld4r {v16.8H, v17.8H, v18.8H, v19.8H}, [x0], #8
> + ld4r {v20.8H, v21.8H, v22.8H, v23.8H}, [x0]
Instead of ld4r like this, would it make things better (or kept similar,
or worse?) if this just would be a regular load, and you'd change the
smlal instructions to take the third operand in the form of e.g. v0.h[0]?
It'd save one instruction when loading the filter coefficients here, but
I'm not sure if there's any other difference...
> +6:
> + mov v3.16B, v1.16B // initialize accumulator part 1 with dithering value
> + mov v4.16B, v2.16B // initialize accumulator part 2 with dithering value
> +
> + ld1 {v24.8H}, [x5], #16 // load 8x16-bit values for src[j + 0][i + {0..7}]
> + ld1 {v25.8H}, [x6], #16 // load 8x16-bit values for src[j + 1][i + {0..7}]
> + ld1 {v26.8H}, [x7], #16 // load 8x16-bit values for src[j + 2][i + {0..7}]
> + ld1 {v27.8H}, [x9], #16 // load 8x16-bit values for src[j + 3][i + {0..7}]
> + ld1 {v28.8H}, [x10], #16 // load 8x16-bit values for src[j + 4][i + {0..7}]
> + ld1 {v29.8H}, [x11], #16 // load 8x16-bit values for src[j + 5][i + {0..7}]
> + ld1 {v30.8H}, [x12], #16 // load 8x16-bit values for src[j + 6][i + {0..7}]
> + ld1 {v31.8H}, [x13], #16 // load 8x16-bit values for src[j + 7][i + {0..7}]
> +
> + smlal v3.4S, v24.4H, v16.4H // val0 += src[0][i + {0..3}] * filter[0]
> + smlal2 v4.4S, v24.8H, v16.8H // val1 += src[0][i + {4..7}] * filter[0]
> + smlal v3.4S, v25.4H, v17.4H // val0 += src[1][i + {0..3}] * filter[1]
> + smlal2 v4.4S, v25.8H, v17.8H // val1 += src[1][i + {4..7}] * filter[1]
> + smlal v3.4S, v26.4H, v18.4H // val0 += src[2][i + {0..3}] * filter[2]
> + smlal2 v4.4S, v26.8H, v18.8H // val1 += src[2][i + {4..7}] * filter[2]
> + smlal v3.4S, v27.4H, v19.4H // val0 += src[3][i + {0..3}] * filter[3]
> + smlal2 v4.4S, v27.8H, v19.8H // val1 += src[3][i + {4..7}] * filter[3]
> + smlal v3.4S, v28.4H, v20.4H // val0 += src[4][i + {0..3}] * filter[4]
> + smlal2 v4.4S, v28.8H, v20.8H // val1 += src[4][i + {4..7}] * filter[4]
> + smlal v3.4S, v29.4H, v21.4H // val0 += src[5][i + {0..3}] * filter[5]
> + smlal2 v4.4S, v29.8H, v21.8H // val1 += src[5][i + {4..7}] * filter[5]
> + smlal v3.4S, v30.4H, v22.4H // val0 += src[6][i + {0..3}] * filter[6]
> + smlal2 v4.4S, v30.8H, v22.8H // val1 += src[6][i + {4..7}] * filter[6]
> + smlal v3.4S, v31.4H, v23.4H // val0 += src[7][i + {0..3}] * filter[7]
> + smlal2 v4.4S, v31.8H, v23.8H // val1 += src[7][i + {4..7}] * filter[7]
> +
> + sqshrun v3.4h, v3.4s, #16 // clip16(val0>>16)
> + sqshrun2 v3.8h, v4.4s, #16 // clip16(val1>>16)
> + uqshrn v3.8b, v3.8h, #3 // clip8(val>>19)
> + st1 {v3.8b}, [x3], #8 // write to destination
> + subs w4, w4, #8 // dstW -= 8
> + b.gt 6b // loop until width consumed
For code like this, I'd usually prefer to move the subs instruction
somewhere else earlier, e.g. between the uqshrn and st1 (where there's a
stall anyway), which moves setting of the condition codes from the subs
further away from the b.gt that uses it.
> + ret
> +
> +7: // fs=4
> + ldp x5, x6, [x2] // load 2 pointers: src[j ] and src[j+1]
> + ldp x7, x9, [x2, #16] // load 2 pointers: src[j+2] and src[j+3]
> +
> + // load 4x16-bit values for filter[j], where j=0..3 and replicated across lanes
> + ld4r {v16.8H, v17.8H, v18.8H, v19.8H}, [x0]
> +8:
> + mov v3.16B, v1.16B // initialize accumulator part 1 with dithering value
> + mov v4.16B, v2.16B // initialize accumulator part 2 with dithering value
> +
> + ld1 {v24.8H}, [x5], #16 // load 8x16-bit values for src[j + 0][i + {0..7}]
> + ld1 {v25.8H}, [x6], #16 // load 8x16-bit values for src[j + 1][i + {0..7}]
> + ld1 {v26.8H}, [x7], #16 // load 8x16-bit values for src[j + 2][i + {0..7}]
> + ld1 {v27.8H}, [x9], #16 // load 8x16-bit values for src[j + 3][i + {0..7}]
> +
> + smlal v3.4S, v24.4H, v16.4H // val0 += src[0][i + {0..3}] * filter[0]
> + smlal2 v4.4S, v24.8H, v16.8H // val1 += src[0][i + {4..7}] * filter[0]
> + smlal v3.4S, v25.4H, v17.4H // val0 += src[1][i + {0..3}] * filter[1]
> + smlal2 v4.4S, v25.8H, v17.8H // val1 += src[1][i + {4..7}] * filter[1]
> + smlal v3.4S, v26.4H, v18.4H // val0 += src[2][i + {0..3}] * filter[2]
> + smlal2 v4.4S, v26.8H, v18.8H // val1 += src[2][i + {4..7}] * filter[2]
> + smlal v3.4S, v27.4H, v19.4H // val0 += src[3][i + {0..3}] * filter[3]
> + smlal2 v4.4S, v27.8H, v19.8H // val1 += src[3][i + {4..7}] * filter[3]
> +
> + sqshrun v3.4h, v3.4s, #16 // clip16(val0>>16)
> + sqshrun2 v3.8h, v4.4s, #16 // clip16(val1>>16)
> + uqshrn v3.8b, v3.8h, #3 // clip8(val>>19)
> + st1 {v3.8b}, [x3], #8 // write to destination
> + subs w4, w4, #8 // dstW -= 8
> + b.gt 8b // loop until width consumed
> + ret
> +
> +9: // fs=2
> + ldp x5, x6, [x2] // load 2 pointers: src[j ] and src[j+1]
> +
> + // load 2x16-bit values for filter[j], where j=0..1 and replicated across lanes
> + ld2r {v16.8H, v17.8H}, [x0]
> +10:
> + mov v3.16B, v1.16B // initialize accumulator part 1 with dithering value
> + mov v4.16B, v2.16B // initialize accumulator part 2 with dithering value
> +
> + ld1 {v24.8H}, [x5], #16 // load 8x16-bit values for src[j + 0][i + {0..7}]
> + ld1 {v25.8H}, [x6], #16 // load 8x16-bit values for src[j + 1][i + {0..7}]
> +
> + smlal v3.4S, v24.4H, v16.4H // val0 += src[0][i + {0..3}] * filter[0]
> + smlal2 v4.4S, v24.8H, v16.8H // val1 += src[0][i + {4..7}] * filter[0]
> + smlal v3.4S, v25.4H, v17.4H // val0 += src[1][i + {0..3}] * filter[1]
> + smlal2 v4.4S, v25.8H, v17.8H // val1 += src[1][i + {4..7}] * filter[1]
> +
> + sqshrun v3.4h, v3.4s, #16 // clip16(val0>>16)
> + sqshrun2 v3.8h, v4.4s, #16 // clip16(val1>>16)
> + uqshrn v3.8b, v3.8h, #3 // clip8(val>>19)
> + st1 {v3.8b}, [x3], #8 // write to destination
> + subs w4, w4, #8 // dstW -= 8
> + b.gt 10b // loop until width consumed
> + ret
> +endfunc
> +function ff_yuv2plane1_8_neon, export=1
Nitpick: It'd be nice with an empty line before this new function
> + ld1 {v0.8B}, [x5] // load 8x8-bit dither
> + cbz w6, 1f // check if offsetting present
> + ext v0.8B, v0.8B, v0.8B, #3 // honor offsetting which can be 0 or 3 only
> +1: uxtl v0.8H, v0.8B // extend dither to 16-bit
> +
> +
> +2:
> + ld1 {v5.8H}, [x0], #16 // read 8x16-bit @ src[j ][i + {0..7}]: A,B,C,D,E,F,G,H
> + add v3.8H, v0.8H, v5.8H
> +
> + uqshrn v3.8b, v3.8h, #7 // clip8(val>>7)
> + st1 {v3.8b}, [x1], #8 // write to destination
> + subs w2, w2, #8 // dstW -= 8
> + b.gt 2b // loop until width consumed
> + ret
> endfunc
Same thing about placement of subs here, I'd prefer to move it up a line
(or maybe even to directly after the ld1?).
> diff --git a/libswscale/aarch64/swscale.c b/libswscale/aarch64/swscale.c
> index 2ea4ccb3a6..0cc821bf11 100644
> --- a/libswscale/aarch64/swscale.c
> +++ b/libswscale/aarch64/swscale.c
> @@ -40,6 +40,12 @@ ALL_SCALE_FUNCS(neon);
> void ff_yuv2planeX_8_neon(const int16_t *filter, int filterSize,
> const int16_t **src, uint8_t *dest, int dstW,
> const uint8_t *dither, int offset);
> +void ff_yuv2plane1_8_neon(
> + const int16_t *src,
> + uint8_t *dest,
> + int dstW,
> + const uint8_t *dither,
> + int offset);
>
> #define ASSIGN_SCALE_FUNC2(hscalefn, filtersize, opt) do { \
> if (c->srcBpc == 8 && c->dstBpc <= 14) { \
> @@ -56,6 +62,11 @@ void ff_yuv2planeX_8_neon(const int16_t *filter, int filterSize,
> ASSIGN_SCALE_FUNC2(hscalefn, X8, opt); \
> break; \
> }
> +#define ASSIGN_VSCALE_FUNC(vscalefn, opt1) \
> + switch(c->dstBpc){ \
> + case 8: vscalefn = ff_yuv2plane1_8_ ## opt1; break; \
> + default: break; \
> + }
Why "opt1" and not just "opt"? Please add spaces between parentheses and
breaces.
Overall this patch looks quite straightforward without anything
significant to add, but I'd like to have the tests extended to cover the
new function and the existing tests extended to cover the filtersize=2
case too.
// Martin
More information about the ffmpeg-devel
mailing list