October 2024 - ffmpeg-cvslog - lists.ffmpeg.org

avcodec/vvc: Don't use large array on stack
by Zhao Zhili 01 Oct '24

01 Oct '24

ffmpeg | branch: master | Zhao Zhili <zhilizhao(a)tencent.com> | Sun Sep 22 11:39:14 2024 +0800| [240c16bbc6ef723139793fea6e94516dfb058db9] | committer: Nuo Mi avcodec/vvc: Don't use large array on stack tmp_array in dmvr_hv takes 33024 bytes on stack, which can be dangerous. > http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=240c16bbc6ef72313… --- libavcodec/vvc/inter_template.c | 33 ++++++++++++++++++--------------- 1 file changed, 18 insertions(+), 15 deletions(-) diff --git a/libavcodec/vvc/inter_template.c b/libavcodec/vvc/inter_template.c index c073a73e76..aee4994c17 100644 --- a/libavcodec/vvc/inter_template.c +++ b/libavcodec/vvc/inter_template.c @@ -472,6 +472,9 @@ static void FUNC(apply_bdof)(uint8_t *_dst, const ptrdiff_t _dst_stride, const i (filter[0] * src[x] + \ filter[1] * src[x + stride]) +#define DMVR_FILTER2(filter, src0, src1) \ + (filter[0] * src0 + filter[1] * src1) + //8.5.3.2.2 Luma sample bilinear interpolation process static void FUNC(dmvr)(int16_t *dst, const uint8_t *_src, const ptrdiff_t _src_stride, const int height, const intptr_t mx, const intptr_t my, const int width) @@ -541,31 +544,31 @@ static void FUNC(dmvr_v)(int16_t *dst, const uint8_t *_src, const ptrdiff_t _src static void FUNC(dmvr_hv)(int16_t *dst, const uint8_t *_src, const ptrdiff_t _src_stride, const int height, const intptr_t mx, const intptr_t my, const int width) { - int16_t tmp_array[(MAX_PB_SIZE + BILINEAR_EXTRA) * MAX_PB_SIZE]; - int16_t *tmp = tmp_array; + int16_t tmp_array[MAX_PB_SIZE * 2]; + int16_t *tmp0 = tmp_array; + int16_t *tmp1 = tmp_array + MAX_PB_SIZE; const pixel *src = (const pixel*)_src; const ptrdiff_t src_stride = _src_stride / sizeof(pixel); - const int8_t *filter = ff_vvc_inter_luma_dmvr_filters[mx]; + const int8_t *filter_x = ff_vvc_inter_luma_dmvr_filters[mx]; + const int8_t *filter_y = ff_vvc_inter_luma_dmvr_filters[my]; const int shift1 = BIT_DEPTH - 6; const int offset1 = 1 << (shift1 - 1); const int shift2 = 4; const int offset2 = 1 << (shift2 - 1); src -= BILINEAR_EXTRA_BEFORE * src_stride; - for (int y = 0; y < height + BILINEAR_EXTRA; y++) { - for (int x = 0; x < width; x++) - tmp[x] = (DMVR_FILTER(src, 1) + offset1) >> shift1; - src += src_stride; - tmp += MAX_PB_SIZE; - } + for (int x = 0; x < width; x++) + tmp0[x] = (DMVR_FILTER2(filter_x, src[x], src[x + 1]) + offset1) >> shift1; + src += src_stride; - tmp = tmp_array + BILINEAR_EXTRA_BEFORE * MAX_PB_SIZE; - filter = ff_vvc_inter_luma_dmvr_filters[my]; - for (int y = 0; y < height; y++) { - for (int x = 0; x < width; x++) - dst[x] = (DMVR_FILTER(tmp, MAX_PB_SIZE) + offset2) >> shift2; - tmp += MAX_PB_SIZE; + for (int y = 1; y < height + BILINEAR_EXTRA; y++) { + for (int x = 0; x < width; x++) { + tmp1[x] = (DMVR_FILTER2(filter_x, src[x], src[x + 1]) + offset1) >> shift1; + dst[x] = (DMVR_FILTER2(filter_y, tmp0[x], tmp1[x]) + offset2) >> shift2; + } + src += src_stride; dst += MAX_PB_SIZE; + FFSWAP(int16_t *, tmp0, tmp1); } }

1 0

aarch64/vvc: Add dmvr
by Zhao Zhili 01 Oct '24

01 Oct '24

ffmpeg | branch: master | Zhao Zhili <zhilizhao(a)tencent.com> | Sun Sep 29 20:02:48 2024 +0800| [5988a2729bc21ade237a9d0f9bb1112f39870f56] | committer: Nuo Mi aarch64/vvc: Add dmvr dmvr_8_12x20_c: 1.5 ( 1.00x) dmvr_8_12x20_neon: 0.2 ( 6.56x) dmvr_8_20x12_c: 1.0 ( 1.00x) dmvr_8_20x12_neon: 0.2 ( 4.33x) dmvr_8_20x20_c: 1.7 ( 1.00x) dmvr_8_20x20_neon: 0.5 ( 3.63x) dmvr_12_12x20_c: 2.2 ( 1.00x) dmvr_12_12x20_neon: 0.5 ( 4.68x) dmvr_12_20x12_c: 2.0 ( 1.00x) dmvr_12_20x12_neon: 0.5 ( 4.16x) dmvr_12_20x20_c: 3.7 ( 1.00x) dmvr_12_20x20_neon: 0.7 ( 5.14x) Signed-off-by: Zhao Zhili <zhilizhao(a)tencent.com> > http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=5988a2729bc21ade2… --- libavcodec/aarch64/vvc/dsp_init.c | 4 ++ libavcodec/aarch64/vvc/inter.S | 87 ++++++++++++++++++++++++++++++++++++++- 2 files changed, 90 insertions(+), 1 deletion(-) diff --git a/libavcodec/aarch64/vvc/dsp_init.c b/libavcodec/aarch64/vvc/dsp_init.c index b5f6f6da93..e47f1ab4cc 100644 --- a/libavcodec/aarch64/vvc/dsp_init.c +++ b/libavcodec/aarch64/vvc/dsp_init.c @@ -90,6 +90,8 @@ W_AVG_FUN(12) const uint8_t *_src, ptrdiff_t _src_stride, int height, \ intptr_t mx, intptr_t my, int width); +DMVR_FUN(, 8) +DMVR_FUN(, 12) DMVR_FUN(hv_, 8) DMVR_FUN(hv_, 10) DMVR_FUN(hv_, 12) @@ -166,6 +168,7 @@ void ff_vvc_dsp_init_aarch64(VVCDSPContext *const c, const int bd) c->inter.avg = ff_vvc_avg_8_neon; c->inter.w_avg = vvc_w_avg_8; + c->inter.dmvr[0][0] = ff_vvc_dmvr_8_neon; c->inter.dmvr[1][1] = ff_vvc_dmvr_hv_8_neon; for (int i = 0; i < FF_ARRAY_ELEMS(c->sao.band_filter); i++) @@ -215,6 +218,7 @@ void ff_vvc_dsp_init_aarch64(VVCDSPContext *const c, const int bd) } else if (bd == 12) { c->inter.avg = ff_vvc_avg_12_neon; c->inter.w_avg = vvc_w_avg_12; + c->inter.dmvr[0][0] = ff_vvc_dmvr_12_neon; c->inter.dmvr[1][1] = ff_vvc_dmvr_hv_12_neon; c->alf.filter[LUMA] = alf_filter_luma_12_neon; diff --git a/libavcodec/aarch64/vvc/inter.S b/libavcodec/aarch64/vvc/inter.S index 4fc8def133..b6b079b569 100644 --- a/libavcodec/aarch64/vvc/inter.S +++ b/libavcodec/aarch64/vvc/inter.S @@ -235,7 +235,7 @@ vvc_avg w_avg, 12 * x5: intptr_t my * w6: int width */ -function ff_vvc_dmvr_hv_8_neon, export=1 +function ff_vvc_dmvr_8_neon, export=1 dst .req x0 src .req x1 src_stride .req x2 @@ -243,6 +243,91 @@ function ff_vvc_dmvr_hv_8_neon, export=1 mx .req x4 my .req x5 width .req w6 + + sxtw x6, w6 + mov x7, #(VVC_MAX_PB_SIZE * 2 + 8) + cmp width, #16 + sub src_stride, src_stride, x6 + cset w15, gt // width > 16 + movi v16.8h, #2 // DMVR_SHIFT + sub x7, x7, x6, lsl #1 +1: + cbz w15, 2f + ldr q0, [src], #16 + uxtl v1.8h, v0.8b + uxtl2 v2.8h, v0.16b + ushl v1.8h, v1.8h, v16.8h + ushl v2.8h, v2.8h, v16.8h + stp q1, q2, [dst], #32 + b 3f +2: + ldr d0, [src], #8 + uxtl v1.8h, v0.8b + ushl v1.8h, v1.8h, v16.8h + str q1, [dst], #16 +3: + subs height, height, #1 + ldr s3, [src], #4 + uxtl v4.8h, v3.8b + ushl v4.4h, v4.4h, v16.4h + st1 {v4.4h}, [dst], x7 + + add src, src, src_stride + b.ne 1b + + ret +endfunc + +function ff_vvc_dmvr_12_neon, export=1 + sxtw x6, w6 + mov x7, #(VVC_MAX_PB_SIZE * 2 + 8) + cmp width, #16 + sub src_stride, src_stride, x6, lsl #1 + cset w15, gt // width > 16 + movi v16.8h, #2 // offset4 + sub x7, x7, x6, lsl #1 +1: + cbz w15, 2f + ldp q0, q1, [src], #32 + uaddl v2.4s, v0.4h, v16.4h + uaddl2 v3.4s, v0.8h, v16.8h + uaddl v4.4s, v1.4h, v16.4h + uaddl2 v5.4s, v1.8h, v16.8h + ushr v2.4s, v2.4s, #2 + ushr v3.4s, v3.4s, #2 + ushr v4.4s, v4.4s, #2 + ushr v5.4s, v5.4s, #2 + uqxtn v2.4h, v2.4s + uqxtn2 v2.8h, v3.4s + uqxtn v4.4h, v4.4s + uqxtn2 v4.8h, v5.4s + + stp q2, q4, [dst], #32 + b 3f +2: + ldr q0, [src], #16 + uaddl v2.4s, v0.4h, v16.4h + uaddl2 v3.4s, v0.8h, v16.8h + ushr v2.4s, v2.4s, #2 + ushr v3.4s, v3.4s, #2 + uqxtn v2.4h, v2.4s + uqxtn2 v2.8h, v3.4s + str q2, [dst], #16 +3: + subs height, height, #1 + ldr d0, [src], #8 + uaddl v3.4s, v0.4h, v16.4h + ushr v3.4s, v3.4s, #2 + uqxtn v3.4h, v3.4s + st1 {v3.4h}, [dst], x7 + + add src, src, src_stride + b.ne 1b + + ret +endfunc + +function ff_vvc_dmvr_hv_8_neon, export=1 tmp0 .req x7 tmp1 .req x8

1 0

aarch64/vvc: Add dmvr_hv
by Zhao Zhili 01 Oct '24

01 Oct '24

ffmpeg | branch: master | Zhao Zhili <zhilizhao(a)tencent.com> | Sun Sep 29 20:02:43 2024 +0800| [bcd65ebd8fcdc4bf81db8a36bac0d7e4ddab0449] | committer: Nuo Mi aarch64/vvc: Add dmvr_hv dmvr_hv_8_12x20_c: 8.0 ( 1.00x) dmvr_hv_8_12x20_neon: 1.2 ( 6.62x) dmvr_hv_8_20x12_c: 8.0 ( 1.00x) dmvr_hv_8_20x12_neon: 0.9 ( 8.37x) dmvr_hv_8_20x20_c: 12.9 ( 1.00x) dmvr_hv_8_20x20_neon: 1.7 ( 7.62x) dmvr_hv_10_12x20_c: 7.0 ( 1.00x) dmvr_hv_10_12x20_neon: 1.7 ( 4.09x) dmvr_hv_10_20x12_c: 7.0 ( 1.00x) dmvr_hv_10_20x12_neon: 1.7 ( 4.09x) dmvr_hv_10_20x20_c: 11.2 ( 1.00x) dmvr_hv_10_20x20_neon: 2.7 ( 4.15x) dmvr_hv_12_12x20_c: 6.5 ( 1.00x) dmvr_hv_12_12x20_neon: 1.7 ( 3.79x) dmvr_hv_12_20x12_c: 6.5 ( 1.00x) dmvr_hv_12_20x12_neon: 1.7 ( 3.79x) dmvr_hv_12_20x20_c: 10.2 ( 1.00x) dmvr_hv_12_20x20_neon: 2.2 ( 4.64x) Signed-off-by: Zhao Zhili <zhilizhao(a)tencent.com> > http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=bcd65ebd8fcdc4bf8… --- libavcodec/aarch64/vvc/dsp_init.c | 12 ++ libavcodec/aarch64/vvc/inter.S | 307 ++++++++++++++++++++++++++++++++++++++ 2 files changed, 319 insertions(+) diff --git a/libavcodec/aarch64/vvc/dsp_init.c b/libavcodec/aarch64/vvc/dsp_init.c index ebe58a2ba5..b5f6f6da93 100644 --- a/libavcodec/aarch64/vvc/dsp_init.c +++ b/libavcodec/aarch64/vvc/dsp_init.c @@ -85,6 +85,15 @@ W_AVG_FUN(8) W_AVG_FUN(10) W_AVG_FUN(12) +#define DMVR_FUN(fn, bd) \ + void ff_vvc_dmvr_ ## fn ## bd ## _neon(int16_t *dst, \ + const uint8_t *_src, ptrdiff_t _src_stride, int height, \ + intptr_t mx, intptr_t my, int width); + +DMVR_FUN(hv_, 8) +DMVR_FUN(hv_, 10) +DMVR_FUN(hv_, 12) + void ff_vvc_dsp_init_aarch64(VVCDSPContext *const c, const int bd) { int cpu_flags = av_get_cpu_flags(); @@ -157,6 +166,7 @@ void ff_vvc_dsp_init_aarch64(VVCDSPContext *const c, const int bd) c->inter.avg = ff_vvc_avg_8_neon; c->inter.w_avg = vvc_w_avg_8; + c->inter.dmvr[1][1] = ff_vvc_dmvr_hv_8_neon; for (int i = 0; i < FF_ARRAY_ELEMS(c->sao.band_filter); i++) c->sao.band_filter[i] = ff_h26x_sao_band_filter_8x8_8_neon; @@ -198,12 +208,14 @@ void ff_vvc_dsp_init_aarch64(VVCDSPContext *const c, const int bd) } else if (bd == 10) { c->inter.avg = ff_vvc_avg_10_neon; c->inter.w_avg = vvc_w_avg_10; + c->inter.dmvr[1][1] = ff_vvc_dmvr_hv_10_neon; c->alf.filter[LUMA] = alf_filter_luma_10_neon; c->alf.filter[CHROMA] = alf_filter_chroma_10_neon; } else if (bd == 12) { c->inter.avg = ff_vvc_avg_12_neon; c->inter.w_avg = vvc_w_avg_12; + c->inter.dmvr[1][1] = ff_vvc_dmvr_hv_12_neon; c->alf.filter[LUMA] = alf_filter_luma_12_neon; c->alf.filter[CHROMA] = alf_filter_chroma_12_neon; diff --git a/libavcodec/aarch64/vvc/inter.S b/libavcodec/aarch64/vvc/inter.S index c4c6ab1a72..4fc8def133 100644 --- a/libavcodec/aarch64/vvc/inter.S +++ b/libavcodec/aarch64/vvc/inter.S @@ -226,3 +226,310 @@ vvc_avg avg, 12 vvc_avg w_avg, 8 vvc_avg w_avg, 10 vvc_avg w_avg, 12 + +/* x0: int16_t *dst + * x1: const uint8_t *_src + * x2: ptrdiff_t _src_stride + * w3: int height + * x4: intptr_t mx + * x5: intptr_t my + * w6: int width + */ +function ff_vvc_dmvr_hv_8_neon, export=1 + dst .req x0 + src .req x1 + src_stride .req x2 + height .req w3 + mx .req x4 + my .req x5 + width .req w6 + tmp0 .req x7 + tmp1 .req x8 + + sub sp, sp, #(VVC_MAX_PB_SIZE * 4) + + movrel x9, X(ff_vvc_inter_luma_dmvr_filters) + add x12, x9, mx, lsl #1 + ldrb w10, [x12] + ldrb w11, [x12, #1] + mov tmp0, sp + add tmp1, tmp0, #(VVC_MAX_PB_SIZE * 2) + // We know the value are positive + dup v0.8h, w10 // filter_x[0] + dup v1.8h, w11 // filter_x[1] + + add x12, x9, my, lsl #1 + ldrb w10, [x12] + ldrb w11, [x12, #1] + sxtw x6, w6 + movi v30.8h, #(1 << (8 - 7)) // offset1 + movi v31.8h, #8 // offset2 + dup v2.8h, w10 // filter_y[0] + dup v3.8h, w11 // filter_y[1] + + // Valid value for width can only be 8 + 4, 16 + 4 + cmp width, #16 + mov w10, #0 // start filter_y or not + add height, height, #1 + sub dst, dst, #(VVC_MAX_PB_SIZE * 2) + sub src_stride, src_stride, x6 + cset w15, gt // width > 16 +1: + mov x12, tmp0 + mov x13, tmp1 + mov x14, dst + cbz w15, 2f + + // width > 16 + ldur q5, [src, #1] + ldr q4, [src], #16 + uxtl v7.8h, v5.8b + uxtl2 v17.8h, v5.16b + uxtl v6.8h, v4.8b + uxtl2 v16.8h, v4.16b + mul v6.8h, v6.8h, v0.8h + mul v16.8h, v16.8h, v0.8h + mla v6.8h, v7.8h, v1.8h + mla v16.8h, v17.8h, v1.8h + add v6.8h, v6.8h, v30.8h + add v16.8h, v16.8h, v30.8h + ushr v6.8h, v6.8h, #(8 - 6) + ushr v7.8h, v16.8h, #(8 - 6) + stp q6, q7, [x13], #32 + + cbz w10, 3f + + ldp q16, q17, [x12], #32 + mul v16.8h, v16.8h, v2.8h + mul v17.8h, v17.8h, v2.8h + mla v16.8h, v6.8h, v3.8h + mla v17.8h, v7.8h, v3.8h + add v16.8h, v16.8h, v31.8h + add v17.8h, v17.8h, v31.8h + ushr v16.8h, v16.8h, #4 + ushr v17.8h, v17.8h, #4 + stp q16, q17, [x14], #32 + b 3f +2: + // width > 8 + ldur d5, [src, #1] + ldr d4, [src], #8 + uxtl v7.8h, v5.8b + uxtl v6.8h, v4.8b + mul v6.8h, v6.8h, v0.8h + mla v6.8h, v7.8h, v1.8h + add v6.8h, v6.8h, v30.8h + ushr v6.8h, v6.8h, #(8 - 6) + str q6, [x13], #16 + + cbz w10, 3f + + ldr q16, [x12], #16 + mul v16.8h, v16.8h, v2.8h + mla v16.8h, v6.8h, v3.8h + add v16.8h, v16.8h, v31.8h + ushr v16.8h, v16.8h, #4 + str q16, [x14], #16 +3: + ldr s5, [src, #1] + ldr s4, [src], #4 + uxtl v7.8h, v5.8b + uxtl v6.8h, v4.8b + mul v6.4h, v6.4h, v0.4h + mla v6.4h, v7.4h, v1.4h + add v6.4h, v6.4h, v30.4h + ushr v6.4h, v6.4h, #(8 - 6) + str d6, [x13], #8 + + cbz w10, 4f + + ldr d16, [x12], #8 + mul v16.4h, v16.4h, v2.4h + mla v16.4h, v6.4h, v3.4h + add v16.4h, v16.4h, v31.4h + ushr v16.4h, v16.4h, #4 + str d16, [x14], #8 +4: + subs height, height, #1 + mov w10, #1 + add src, src, src_stride + add dst, dst, #(VVC_MAX_PB_SIZE * 2) + eor tmp0, tmp0, tmp1 + eor tmp1, tmp0, tmp1 + eor tmp0, tmp0, tmp1 + b.ne 1b + + add sp, sp, #(VVC_MAX_PB_SIZE * 4) + ret +endfunc + +function ff_vvc_dmvr_hv_12_neon, export=1 + movi v29.4s, #(12 - 6) + movi v30.4s, #(1 << (12 - 7)) // offset1 + b 0f +endfunc + +function ff_vvc_dmvr_hv_10_neon, export=1 + movi v29.4s, #(10 - 6) + movi v30.4s, #(1 << (10 - 7)) // offset1 +0: + movi v31.4s, #8 // offset2 + neg v29.4s, v29.4s + + sub sp, sp, #(VVC_MAX_PB_SIZE * 4) + + movrel x9, X(ff_vvc_inter_luma_dmvr_filters) + add x12, x9, mx, lsl #1 + ldrb w10, [x12] + ldrb w11, [x12, #1] + mov tmp0, sp + add tmp1, tmp0, #(VVC_MAX_PB_SIZE * 2) + // We know the value are positive + dup v0.8h, w10 // filter_x[0] + dup v1.8h, w11 // filter_x[1] + + add x12, x9, my, lsl #1 + ldrb w10, [x12] + ldrb w11, [x12, #1] + sxtw x6, w6 + dup v2.8h, w10 // filter_y[0] + dup v3.8h, w11 // filter_y[1] + + // Valid value for width can only be 8 + 4, 16 + 4 + cmp width, #16 + mov w10, #0 // start filter_y or not + add height, height, #1 + sub dst, dst, #(VVC_MAX_PB_SIZE * 2) + sub src_stride, src_stride, x6, lsl #1 + cset w15, gt // width > 16 +1: + mov x12, tmp0 + mov x13, tmp1 + mov x14, dst + cbz w15, 2f + + // width > 16 + add x16, src, #2 + ldp q6, q16, [src], #32 + ldp q7, q17, [x16] + umull v4.4s, v6.4h, v0.4h + umull2 v5.4s, v6.8h, v0.8h + umull v18.4s, v16.4h, v0.4h + umull2 v19.4s, v16.8h, v0.8h + umlal v4.4s, v7.4h, v1.4h + umlal2 v5.4s, v7.8h, v1.8h + umlal v18.4s, v17.4h, v1.4h + umlal2 v19.4s, v17.8h, v1.8h + + add v4.4s, v4.4s, v30.4s + add v5.4s, v5.4s, v30.4s + add v18.4s, v18.4s, v30.4s + add v19.4s, v19.4s, v30.4s + ushl v4.4s, v4.4s, v29.4s + ushl v5.4s, v5.4s, v29.4s + ushl v18.4s, v18.4s, v29.4s + ushl v19.4s, v19.4s, v29.4s + uqxtn v6.4h, v4.4s + uqxtn2 v6.8h, v5.4s + uqxtn v7.4h, v18.4s + uqxtn2 v7.8h, v19.4s + stp q6, q7, [x13], #32 + + cbz w10, 3f + + ldp q4, q5, [x12], #32 + umull v17.4s, v4.4h, v2.4h + umull2 v18.4s, v4.8h, v2.8h + umull v19.4s, v5.4h, v2.4h + umull2 v20.4s, v5.8h, v2.8h + umlal v17.4s, v6.4h, v3.4h + umlal2 v18.4s, v6.8h, v3.8h + umlal v19.4s, v7.4h, v3.4h + umlal2 v20.4s, v7.8h, v3.8h + add v17.4s, v17.4s, v31.4s + add v18.4s, v18.4s, v31.4s + add v19.4s, v19.4s, v31.4s + add v20.4s, v20.4s, v31.4s + ushr v17.4s, v17.4s, #4 + ushr v18.4s, v18.4s, #4 + ushr v19.4s, v19.4s, #4 + ushr v20.4s, v20.4s, #4 + uqxtn v6.4h, v17.4s + uqxtn2 v6.8h, v18.4s + uqxtn v7.4h, v19.4s + uqxtn2 v7.8h, v20.4s + stp q6, q7, [x14], #32 + b 3f +2: + // width > 8 + ldur q7, [src, #2] + ldr q6, [src], #16 + umull v4.4s, v6.4h, v0.4h + umull2 v5.4s, v6.8h, v0.8h + umlal v4.4s, v7.4h, v1.4h + umlal2 v5.4s, v7.8h, v1.8h + + add v4.4s, v4.4s, v30.4s + add v5.4s, v5.4s, v30.4s + ushl v4.4s, v4.4s, v29.4s + ushl v5.4s, v5.4s, v29.4s + uqxtn v6.4h, v4.4s + uqxtn2 v6.8h, v5.4s + str q6, [x13], #16 + + cbz w10, 3f + + ldr q16, [x12], #16 + umull v17.4s, v16.4h, v2.4h + umull2 v18.4s, v16.8h, v2.8h + umlal v17.4s, v6.4h, v3.4h + umlal2 v18.4s, v6.8h, v3.8h + add v17.4s, v17.4s, v31.4s + add v18.4s, v18.4s, v31.4s + ushr v17.4s, v17.4s, #4 + ushr v18.4s, v18.4s, #4 + uqxtn v16.4h, v17.4s + uqxtn2 v16.8h, v18.4s + str q16, [x14], #16 +3: + ldr d7, [src, #2] + ldr d6, [src], #8 + umull v4.4s, v7.4h, v1.4h + umlal v4.4s, v6.4h, v0.4h + add v4.4s, v4.4s, v30.4s + ushl v4.4s, v4.4s, v29.4s + uqxtn v6.4h, v4.4s + str d6, [x13], #8 + + cbz w10, 4f + + ldr d16, [x12], #8 + umull v17.4s, v16.4h, v2.4h + umlal v17.4s, v6.4h, v3.4h + add v17.4s, v17.4s, v31.4s + ushr v17.4s, v17.4s, #4 + uqxtn v16.4h, v17.4s + str d16, [x14], #8 +4: + subs height, height, #1 + mov w10, #1 + add src, src, src_stride + add dst, dst, #(VVC_MAX_PB_SIZE * 2) + eor tmp0, tmp0, tmp1 + eor tmp1, tmp0, tmp1 + eor tmp0, tmp0, tmp1 + b.ne 1b + + add sp, sp, #(VVC_MAX_PB_SIZE * 4) + ret + +.unreq dst +.unreq src +.unreq src_stride +.unreq height +.unreq mx +.unreq my +.unreq width +.unreq tmp0 +.unreq tmp1 +endfunc

1 0

aarch64/vvc: Add w_avg
by Zhao Zhili 01 Oct '24

01 Oct '24

ffmpeg | branch: master | Zhao Zhili <zhilizhao(a)tencent.com> | Sun Sep 29 20:02:27 2024 +0800| [0ba9e8d0d492b8477eabbd8ec0c114bf05b7d7b7] | committer: Nuo Mi aarch64/vvc: Add w_avg w_avg_8_2x2_c: 0.0 ( 0.00x) w_avg_8_2x2_neon: 0.0 ( 0.00x) w_avg_8_4x4_c: 0.2 ( 1.00x) w_avg_8_4x4_neon: 0.0 ( 0.00x) w_avg_8_8x8_c: 1.2 ( 1.00x) w_avg_8_8x8_neon: 0.2 ( 5.00x) w_avg_8_16x16_c: 4.2 ( 1.00x) w_avg_8_16x16_neon: 0.8 ( 5.67x) w_avg_8_32x32_c: 16.2 ( 1.00x) w_avg_8_32x32_neon: 2.5 ( 6.50x) w_avg_8_64x64_c: 64.5 ( 1.00x) w_avg_8_64x64_neon: 9.0 ( 7.17x) w_avg_8_128x128_c: 269.5 ( 1.00x) w_avg_8_128x128_neon: 35.5 ( 7.59x) w_avg_10_2x2_c: 0.2 ( 1.00x) w_avg_10_2x2_neon: 0.2 ( 1.00x) w_avg_10_4x4_c: 0.2 ( 1.00x) w_avg_10_4x4_neon: 0.2 ( 1.00x) w_avg_10_8x8_c: 1.0 ( 1.00x) w_avg_10_8x8_neon: 0.2 ( 4.00x) w_avg_10_16x16_c: 4.2 ( 1.00x) w_avg_10_16x16_neon: 0.8 ( 5.67x) w_avg_10_32x32_c: 16.2 ( 1.00x) w_avg_10_32x32_neon: 2.5 ( 6.50x) w_avg_10_64x64_c: 66.2 ( 1.00x) w_avg_10_64x64_neon: 10.0 ( 6.62x) w_avg_10_128x128_c: 277.8 ( 1.00x) w_avg_10_128x128_neon: 39.8 ( 6.99x) w_avg_12_2x2_c: 0.0 ( 0.00x) w_avg_12_2x2_neon: 0.2 ( 0.00x) w_avg_12_4x4_c: 0.2 ( 1.00x) w_avg_12_4x4_neon: 0.0 ( 0.00x) w_avg_12_8x8_c: 1.2 ( 1.00x) w_avg_12_8x8_neon: 0.5 ( 2.50x) w_avg_12_16x16_c: 4.8 ( 1.00x) w_avg_12_16x16_neon: 0.8 ( 6.33x) w_avg_12_32x32_c: 17.0 ( 1.00x) w_avg_12_32x32_neon: 2.8 ( 6.18x) w_avg_12_64x64_c: 64.0 ( 1.00x) w_avg_12_64x64_neon: 10.0 ( 6.40x) w_avg_12_128x128_c: 269.2 ( 1.00x) w_avg_12_128x128_neon: 42.0 ( 6.41x) Signed-off-by: Zhao Zhili <zhilizhao(a)tencent.com> > http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=0ba9e8d0d492b8477… --- libavcodec/aarch64/vvc/dsp_init.c | 36 ++++++++++++++ libavcodec/aarch64/vvc/inter.S | 99 ++++++++++++++++++++++++++++++++------- 2 files changed, 118 insertions(+), 17 deletions(-) diff --git a/libavcodec/aarch64/vvc/dsp_init.c b/libavcodec/aarch64/vvc/dsp_init.c index ad767d17e2..ebe58a2ba5 100644 --- a/libavcodec/aarch64/vvc/dsp_init.c +++ b/libavcodec/aarch64/vvc/dsp_init.c @@ -52,6 +52,39 @@ void ff_vvc_avg_12_neon(uint8_t *dst, ptrdiff_t dst_stride, const int16_t *src0, const int16_t *src1, int width, int height); +void ff_vvc_w_avg_8_neon(uint8_t *_dst, ptrdiff_t _dst_stride, + const int16_t *src0, const int16_t *src1, + int width, int height, + uintptr_t w0_w1, uintptr_t offset_shift); +void ff_vvc_w_avg_10_neon(uint8_t *_dst, ptrdiff_t _dst_stride, + const int16_t *src0, const int16_t *src1, + int width, int height, + uintptr_t w0_w1, uintptr_t offset_shift); +void ff_vvc_w_avg_12_neon(uint8_t *_dst, ptrdiff_t _dst_stride, + const int16_t *src0, const int16_t *src1, + int width, int height, + uintptr_t w0_w1, uintptr_t offset_shift); +/* When passing arguments to functions, Apple platforms diverge from the ARM64 + * standard ABI for functions that require passing arguments on the stack. To + * simplify portability in the assembly function interface, use a different + * function signature that doesn't require passing arguments on the stack. + */ +#define W_AVG_FUN(bit_depth) \ +static void vvc_w_avg_ ## bit_depth(uint8_t *dst, ptrdiff_t dst_stride, \ + const int16_t *src0, const int16_t *src1, int width, int height, \ + int denom, int w0, int w1, int o0, int o1) \ +{ \ + int shift = denom + FFMAX(3, 15 - bit_depth); \ + int offset = ((o0 + o1) * (1 << (bit_depth - 8)) + 1) * (1 << (shift - 1)); \ + uintptr_t w0_w1 = ((uintptr_t)w0 << 32) | (uint32_t)w1; \ + uintptr_t offset_shift = ((uintptr_t)offset << 32) | (uint32_t)shift; \ + ff_vvc_w_avg_ ## bit_depth ## _neon(dst, dst_stride, src0, src1, width, height, w0_w1, offset_shift); \ +} + +W_AVG_FUN(8) +W_AVG_FUN(10) +W_AVG_FUN(12) + void ff_vvc_dsp_init_aarch64(VVCDSPContext *const c, const int bd) { int cpu_flags = av_get_cpu_flags(); @@ -123,6 +156,7 @@ void ff_vvc_dsp_init_aarch64(VVCDSPContext *const c, const int bd) c->inter.put_uni_w[0][6][0][0] = ff_vvc_put_pel_uni_w_pixels128_8_neon; c->inter.avg = ff_vvc_avg_8_neon; + c->inter.w_avg = vvc_w_avg_8; for (int i = 0; i < FF_ARRAY_ELEMS(c->sao.band_filter); i++) c->sao.band_filter[i] = ff_h26x_sao_band_filter_8x8_8_neon; @@ -163,11 +197,13 @@ void ff_vvc_dsp_init_aarch64(VVCDSPContext *const c, const int bd) } } else if (bd == 10) { c->inter.avg = ff_vvc_avg_10_neon; + c->inter.w_avg = vvc_w_avg_10; c->alf.filter[LUMA] = alf_filter_luma_10_neon; c->alf.filter[CHROMA] = alf_filter_chroma_10_neon; } else if (bd == 12) { c->inter.avg = ff_vvc_avg_12_neon; + c->inter.w_avg = vvc_w_avg_12; c->alf.filter[LUMA] = alf_filter_luma_12_neon; c->alf.filter[CHROMA] = alf_filter_chroma_12_neon; diff --git a/libavcodec/aarch64/vvc/inter.S b/libavcodec/aarch64/vvc/inter.S index 2f69274b86..c4c6ab1a72 100644 --- a/libavcodec/aarch64/vvc/inter.S +++ b/libavcodec/aarch64/vvc/inter.S @@ -22,9 +22,9 @@ #define VVC_MAX_PB_SIZE 128 -.macro vvc_avg, bit_depth +.macro vvc_avg type, bit_depth -.macro vvc_avg_\bit_depth\()_2_4, tap +.macro vvc_\type\()_\bit_depth\()_2_4 tap .if \tap == 2 ldr s0, [src0] ldr s2, [src1] @@ -32,9 +32,19 @@ ldr d0, [src0] ldr d2, [src1] .endif + +.ifc \type, avg saddl v4.4s, v0.4h, v2.4h add v4.4s, v4.4s, v16.4s sqshrn v4.4h, v4.4s, #(15 - \bit_depth) +.else + mov v4.16b, v16.16b + smlal v4.4s, v0.4h, v19.4h + smlal v4.4s, v2.4h, v20.4h + sqshl v4.4s, v4.4s, v22.4s + sqxtn v4.4h, v4.4s +.endif + .if \bit_depth == 8 sqxtun v4.8b, v4.8h .if \tap == 2 @@ -57,7 +67,7 @@ add dst, dst, dst_stride .endm -function ff_vvc_avg_\bit_depth\()_neon, export=1 +function ff_vvc_\type\()_\bit_depth\()_neon, export=1 dst .req x0 dst_stride .req x1 src0 .req x2 @@ -67,42 +77,64 @@ function ff_vvc_avg_\bit_depth\()_neon, export=1 mov x10, #(VVC_MAX_PB_SIZE * 2) cmp width, #8 -.if \bit_depth == 8 - movi v16.4s, #64 -.else -.if \bit_depth == 10 - mov w6, #1023 - movi v16.4s, #16 +.ifc \type, avg + movi v16.4s, #(1 << (14 - \bit_depth)) .else - mov w6, #4095 - movi v16.4s, #4 -.endif + lsr x11, x6, #32 // weight0 + mov w12, w6 // weight1 + lsr x13, x7, #32 // offset + mov w14, w7 // shift + + dup v19.8h, w11 + neg w14, w14 // so we can use sqshl + dup v20.8h, w12 + dup v16.4s, w13 + dup v22.4s, w14 +.endif // avg + + .if \bit_depth >= 10 + // clip pixel + mov w6, #((1 << \bit_depth) - 1) movi v18.8h, #0 dup v17.8h, w6 .endif + b.eq 8f b.hi 16f cmp width, #4 b.eq 4f 2: // width == 2 subs height, height, #1 - vvc_avg_\bit_depth\()_2_4 2 + vvc_\type\()_\bit_depth\()_2_4 2 b.ne 2b b 32f 4: // width == 4 subs height, height, #1 - vvc_avg_\bit_depth\()_2_4 4 + vvc_\type\()_\bit_depth\()_2_4 4 b.ne 4b b 32f 8: // width == 8 ld1 {v0.8h}, [src0], x10 ld1 {v2.8h}, [src1], x10 +.ifc \type, avg saddl v4.4s, v0.4h, v2.4h saddl2 v5.4s, v0.8h, v2.8h add v4.4s, v4.4s, v16.4s add v5.4s, v5.4s, v16.4s sqshrn v4.4h, v4.4s, #(15 - \bit_depth) sqshrn2 v4.8h, v5.4s, #(15 - \bit_depth) +.else + mov v4.16b, v16.16b + mov v5.16b, v16.16b + smlal v4.4s, v0.4h, v19.4h + smlal v4.4s, v2.4h, v20.4h + smlal2 v5.4s, v0.8h, v19.8h + smlal2 v5.4s, v2.8h, v20.8h + sqshl v4.4s, v4.4s, v22.4s + sqshl v5.4s, v5.4s, v22.4s + sqxtn v4.4h, v4.4s + sqxtn2 v4.8h, v5.4s +.endif subs height, height, #1 .if \bit_depth == 8 sqxtun v4.8b, v4.8h @@ -122,6 +154,7 @@ function ff_vvc_avg_\bit_depth\()_neon, export=1 17: ldp q0, q1, [x7], #32 ldp q2, q3, [x8], #32 +.ifc \type, avg saddl v4.4s, v0.4h, v2.4h saddl2 v5.4s, v0.8h, v2.8h saddl v6.4s, v1.4h, v3.4h @@ -134,6 +167,28 @@ function ff_vvc_avg_\bit_depth\()_neon, export=1 sqshrn2 v4.8h, v5.4s, #(15 - \bit_depth) sqshrn v6.4h, v6.4s, #(15 - \bit_depth) sqshrn2 v6.8h, v7.4s, #(15 - \bit_depth) +.else // avg + mov v4.16b, v16.16b + mov v5.16b, v16.16b + mov v6.16b, v16.16b + mov v7.16b, v16.16b + smlal v4.4s, v0.4h, v19.4h + smlal v4.4s, v2.4h, v20.4h + smlal2 v5.4s, v0.8h, v19.8h + smlal2 v5.4s, v2.8h, v20.8h + smlal v6.4s, v1.4h, v19.4h + smlal v6.4s, v3.4h, v20.4h + smlal2 v7.4s, v1.8h, v19.8h + smlal2 v7.4s, v3.8h, v20.8h + sqshl v4.4s, v4.4s, v22.4s + sqshl v5.4s, v5.4s, v22.4s + sqshl v6.4s, v6.4s, v22.4s + sqshl v7.4s, v7.4s, v22.4s + sqxtn v4.4h, v4.4s + sqxtn v6.4h, v6.4s + sqxtn2 v4.8h, v5.4s + sqxtn2 v6.8h, v7.4s +.endif // w_avg subs w6, w6, #16 .if \bit_depth == 8 sqxtun v4.8b, v4.8h @@ -155,9 +210,19 @@ function ff_vvc_avg_\bit_depth\()_neon, export=1 b.ne 16b 32: ret + +.unreq dst +.unreq dst_stride +.unreq src0 +.unreq src1 +.unreq width +.unreq height endfunc .endm -vvc_avg 8 -vvc_avg 10 -vvc_avg 12 +vvc_avg avg, 8 +vvc_avg avg, 10 +vvc_avg avg, 12 +vvc_avg w_avg, 8 +vvc_avg w_avg, 10 +vvc_avg w_avg, 12

1 0