History

lavc/aarch64: new optimization for 8-bit hevc_epel_bi_hv

put_hevc_epel_bi_hv4_8_c: 242.9 put_hevc_epel_bi_hv4_8_i8mm: 68.6 put_hevc_epel_bi_hv6_8_c: 402.4 put_hevc_epel_bi_hv6_8_i8mm: 135.9 put_hevc_epel_bi_hv8_8_c: 636.4 put_hevc_epel_bi_hv8_8_i8mm: 145.6 put_hevc_epel_bi_hv12_8_c: 1363.1 put_hevc_epel_bi_hv12_8_i8mm: 324.1 put_hevc_epel_bi_hv16_8_c: 2222.1 put_hevc_epel_bi_hv16_8_i8mm: 509.1 put_hevc_epel_bi_hv24_8_c: 4793.4 put_hevc_epel_bi_hv24_8_i8mm: 1091.9 put_hevc_epel_bi_hv32_8_c: 8393.9 put_hevc_epel_bi_hv32_8_i8mm: 1720.6 put_hevc_epel_bi_hv48_8_c: 19526.6 put_hevc_epel_bi_hv48_8_i8mm: 4285.9 put_hevc_epel_bi_hv64_8_c: 33915.4 put_hevc_epel_bi_hv64_8_i8mm: 6783.6

Logan Lyu committed 3 years ago

Commit 00290a61 parent 0448f272 files · +335 −0Browse files

Unified · Split

2 files
  1. libavcodec/aarch64/hevcdsp_epel_neon.S +330 −0
  2. libavcodec/aarch64/hevcdsp_init_aarch64.c +5 −0
libavcodec/aarch64/hevcdsp_epel_neon.S+330−0
@@ -3203,6 +3203,336 @@ function ff_hevc_put_hevc_epel_uni_w_hv64_8_neon_i8mm, export=1 · 3202 unmodified lines
32033203 ret
32043204 endfunc
32053205
3206+
3207+function ff_hevc_put_hevc_epel_bi_hv4_8_neon_i8mm, export=1
3208+ add w10, w5, #3
3209+ lsl x10, x10, #7
3210+ sub sp, sp, x10 // tmp_array
3211+ stp x7, x30, [sp, #-48]!
3212+ stp x4, x5, [sp, #16]
3213+ stp x0, x1, [sp, #32]
3214+ add x0, sp, #48
3215+ sub x1, x2, x3
3216+ mov x2, x3
3217+ add w3, w5, #3
3218+ mov x4, x6
3219+ mov x5, x7
3220+ bl X(ff_hevc_put_hevc_epel_h4_8_neon_i8mm)
3221+ ldp x4, x5, [sp, #16]
3222+ ldp x0, x1, [sp, #32]
3223+ ldp x7, x30, [sp], #48
3224+ load_epel_filterh x7, x6
3225+ mov x10, #(MAX_PB_SIZE * 2)
3226+ ld1 {v16.4h}, [sp], x10
3227+ ld1 {v17.4h}, [sp], x10
3228+ ld1 {v18.4h}, [sp], x10
3229+.macro calc src0, src1, src2, src3
3230+ ld1 {\src3\().4h}, [sp], x10
3231+ calc_epelh v4, \src0, \src1, \src2, \src3
3232+ ld1 {v6.4h}, [x4], x10
3233+ sqadd v4.4h, v4.4h, v6.4h
3234+ sqrshrun v4.8b, v4.8h, #7
3235+ subs w5, w5, #1
3236+ st1 {v4.s}[0], [x0], x1
3237+.endm
3238+1: calc_all4
3239+.purgem calc
3240+2: ret
3241+endfunc
3242+
3243+function ff_hevc_put_hevc_epel_bi_hv6_8_neon_i8mm, export=1
3244+ add w10, w5, #3
3245+ lsl x10, x10, #7
3246+ sub sp, sp, x10 // tmp_array
3247+ stp x7, x30, [sp, #-48]!
3248+ stp x4, x5, [sp, #16]
3249+ stp x0, x1, [sp, #32]
3250+ add x0, sp, #48
3251+ sub x1, x2, x3
3252+ mov x2, x3
3253+ add w3, w5, #3
3254+ mov x4, x6
3255+ mov x5, x7
3256+ bl X(ff_hevc_put_hevc_epel_h6_8_neon_i8mm)
3257+ ldp x4, x5, [sp, #16]
3258+ ldp x0, x1, [sp, #32]
3259+ ldp x7, x30, [sp], #48
3260+ load_epel_filterh x7, x6
3261+ sub x1, x1, #4
3262+ mov x10, #(MAX_PB_SIZE * 2)
3263+ ld1 {v16.8h}, [sp], x10
3264+ ld1 {v17.8h}, [sp], x10
3265+ ld1 {v18.8h}, [sp], x10
3266+.macro calc src0, src1, src2, src3
3267+ ld1 {\src3\().8h}, [sp], x10
3268+ calc_epelh v4, \src0, \src1, \src2, \src3
3269+ calc_epelh2 v4, v5, \src0, \src1, \src2, \src3
3270+ ld1 {v6.8h}, [x4], x10
3271+ sqadd v4.8h, v4.8h, v6.8h
3272+ sqrshrun v4.8b, v4.8h, #7
3273+ st1 {v4.s}[0], [x0], #4
3274+ subs w5, w5, #1
3275+ st1 {v4.h}[2], [x0], x1
3276+.endm
3277+1: calc_all4
3278+.purgem calc
3279+2: ret
3280+endfunc
3281+
3282+function ff_hevc_put_hevc_epel_bi_hv8_8_neon_i8mm, export=1
3283+ add w10, w5, #3
3284+ lsl x10, x10, #7
3285+ sub sp, sp, x10 // tmp_array
3286+ stp x7, x30, [sp, #-48]!
3287+ stp x4, x5, [sp, #16]
3288+ stp x0, x1, [sp, #32]
3289+ add x0, sp, #48
3290+ sub x1, x2, x3
3291+ mov x2, x3
3292+ add w3, w5, #3
3293+ mov x4, x6
3294+ mov x5, x7
3295+ bl X(ff_hevc_put_hevc_epel_h8_8_neon_i8mm)
3296+ ldp x4, x5, [sp, #16]
3297+ ldp x0, x1, [sp, #32]
3298+ ldp x7, x30, [sp], #48
3299+ load_epel_filterh x7, x6
3300+ mov x10, #(MAX_PB_SIZE * 2)
3301+ ld1 {v16.8h}, [sp], x10
3302+ ld1 {v17.8h}, [sp], x10
3303+ ld1 {v18.8h}, [sp], x10
3304+.macro calc src0, src1, src2, src3
3305+ ld1 {\src3\().8h}, [sp], x10
3306+ calc_epelh v4, \src0, \src1, \src2, \src3
3307+ calc_epelh2 v4, v5, \src0, \src1, \src2, \src3
3308+ ld1 {v6.8h}, [x4], x10
3309+ sqadd v4.8h, v4.8h, v6.8h
3310+ sqrshrun v4.8b, v4.8h, #7
3311+ subs w5, w5, #1
3312+ st1 {v4.8b}, [x0], x1
3313+.endm
3314+1: calc_all4
3315+.purgem calc
3316+2: ret
3317+endfunc
3318+
3319+function ff_hevc_put_hevc_epel_bi_hv12_8_neon_i8mm, export=1
3320+ add w10, w5, #3
3321+ lsl x10, x10, #7
3322+ sub sp, sp, x10 // tmp_array
3323+ stp x7, x30, [sp, #-48]!
3324+ stp x4, x5, [sp, #16]
3325+ stp x0, x1, [sp, #32]
3326+ add x0, sp, #48
3327+ sub x1, x2, x3
3328+ mov x2, x3
3329+ add w3, w5, #3
3330+ mov x4, x6
3331+ mov x5, x7
3332+ bl X(ff_hevc_put_hevc_epel_h12_8_neon_i8mm)
3333+ ldp x4, x5, [sp, #16]
3334+ ldp x0, x1, [sp, #32]
3335+ ldp x7, x30, [sp], #48
3336+ load_epel_filterh x7, x6
3337+ sub x1, x1, #8
3338+ mov x10, #(MAX_PB_SIZE * 2)
3339+ ld1 {v16.8h, v17.8h}, [sp], x10
3340+ ld1 {v18.8h, v19.8h}, [sp], x10
3341+ ld1 {v20.8h, v21.8h}, [sp], x10
3342+.macro calc src0, src1, src2, src3, src4, src5, src6, src7
3343+ ld1 {\src6\().8h, \src7\().8h}, [sp], x10
3344+ calc_epelh v4, \src0, \src2, \src4, \src6
3345+ calc_epelh2 v4, v5, \src0, \src2, \src4, \src6
3346+ calc_epelh v5, \src1, \src3, \src5, \src7
3347+ ld1 {v6.8h, v7.8h}, [x4], x10
3348+ sqadd v4.8h, v4.8h, v6.8h
3349+ sqadd v5.8h, v5.8h, v7.8h
3350+ sqrshrun v4.8b, v4.8h, #7
3351+ sqrshrun2 v4.16b, v5.8h, #7
3352+ st1 {v4.8b}, [x0], #8
3353+ subs w5, w5, #1
3354+ st1 {v4.s}[2], [x0], x1
3355+.endm
3356+1: calc_all8
3357+.purgem calc
3358+2: ret
3359+endfunc
3360+
3361+function ff_hevc_put_hevc_epel_bi_hv16_8_neon_i8mm, export=1
3362+ add w10, w5, #3
3363+ lsl x10, x10, #7
3364+ sub sp, sp, x10 // tmp_array
3365+ stp x7, x30, [sp, #-48]!
3366+ stp x4, x5, [sp, #16]
3367+ stp x0, x1, [sp, #32]
3368+ add x0, sp, #48
3369+ sub x1, x2, x3
3370+ mov x2, x3
3371+ add w3, w5, #3
3372+ mov x4, x6
3373+ mov x5, x7
3374+ bl X(ff_hevc_put_hevc_epel_h16_8_neon_i8mm)
3375+ ldp x4, x5, [sp, #16]
3376+ ldp x0, x1, [sp, #32]
3377+ ldp x7, x30, [sp], #48
3378+ load_epel_filterh x7, x6
3379+ mov x10, #(MAX_PB_SIZE * 2)
3380+ ld1 {v16.8h, v17.8h}, [sp], x10
3381+ ld1 {v18.8h, v19.8h}, [sp], x10
3382+ ld1 {v20.8h, v21.8h}, [sp], x10
3383+.macro calc src0, src1, src2, src3, src4, src5, src6, src7
3384+ ld1 {\src6\().8h, \src7\().8h}, [sp], x10
3385+ calc_epelh v4, \src0, \src2, \src4, \src6
3386+ calc_epelh2 v4, v5, \src0, \src2, \src4, \src6
3387+ calc_epelh v5, \src1, \src3, \src5, \src7
3388+ calc_epelh2 v5, v6, \src1, \src3, \src5, \src7
3389+ ld1 {v6.8h, v7.8h}, [x4], x10
3390+ sqadd v4.8h, v4.8h, v6.8h
3391+ sqadd v5.8h, v5.8h, v7.8h
3392+ sqrshrun v4.8b, v4.8h, #7
3393+ sqrshrun2 v4.16b, v5.8h, #7
3394+ st1 {v4.16b}, [x0], x1
3395+ subs w5, w5, #1
3396+.endm
3397+1: calc_all8
3398+.purgem calc
3399+2: ret
3400+endfunc
3401+
3402+function ff_hevc_put_hevc_epel_bi_hv24_8_neon_i8mm, export=1
3403+ add w10, w5, #3
3404+ lsl x10, x10, #7
3405+ sub sp, sp, x10 // tmp_array
3406+ stp x7, x30, [sp, #-48]!
3407+ stp x4, x5, [sp, #16]
3408+ stp x0, x1, [sp, #32]
3409+ add x0, sp, #48
3410+ sub x1, x2, x3
3411+ mov x2, x3
3412+ add w3, w5, #3
3413+ mov x4, x6
3414+ mov x5, x7
3415+ bl X(ff_hevc_put_hevc_epel_h24_8_neon_i8mm)
3416+ ldp x4, x5, [sp, #16]
3417+ ldp x0, x1, [sp, #32]
3418+ ldp x7, x30, [sp], #48
3419+ load_epel_filterh x7, x6
3420+ mov x10, #(MAX_PB_SIZE * 2)
3421+ ld1 {v16.8h, v17.8h, v18.8h}, [sp], x10
3422+ ld1 {v19.8h, v20.8h, v21.8h}, [sp], x10
3423+ ld1 {v22.8h, v23.8h, v24.8h}, [sp], x10
3424+.macro calc src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10, src11
3425+ ld1 {\src9\().8h, \src10\().8h, \src11\().8h}, [sp], x10
3426+ calc_epelh v1, \src0, \src3, \src6, \src9
3427+ calc_epelh2 v1, v2, \src0, \src3, \src6, \src9
3428+ calc_epelh v2, \src1, \src4, \src7, \src10
3429+ calc_epelh2 v2, v3, \src1, \src4, \src7, \src10
3430+ calc_epelh v3, \src2, \src5, \src8, \src11
3431+ calc_epelh2 v3, v4, \src2, \src5, \src8, \src11
3432+ ld1 {v4.8h, v5.8h, v6.8h}, [x4], x10
3433+ sqadd v1.8h, v1.8h, v4.8h
3434+ sqadd v2.8h, v2.8h, v5.8h
3435+ sqadd v3.8h, v3.8h, v6.8h
3436+ sqrshrun v1.8b, v1.8h, #7
3437+ sqrshrun v2.8b, v2.8h, #7
3438+ sqrshrun v3.8b, v3.8h, #7
3439+ subs w5, w5, #1
3440+ st1 {v1.8b, v2.8b, v3.8b}, [x0], x1
3441+.endm
3442+1: calc_all12
3443+.purgem calc
3444+2: ret
3445+endfunc
3446+
3447+function ff_hevc_put_hevc_epel_bi_hv32_8_neon_i8mm, export=1
3448+ str d8, [sp, #-16]!
3449+ add w10, w5, #3
3450+ lsl x10, x10, #7
3451+ sub sp, sp, x10 // tmp_array
3452+ stp x7, x30, [sp, #-48]!
3453+ stp x4, x5, [sp, #16]
3454+ stp x0, x1, [sp, #32]
3455+ add x0, sp, #48
3456+ sub x1, x2, x3
3457+ mov x2, x3
3458+ add w3, w5, #3
3459+ mov x4, x6
3460+ mov x5, x7
3461+ bl X(ff_hevc_put_hevc_epel_h32_8_neon_i8mm)
3462+ ldp x4, x5, [sp, #16]
3463+ ldp x0, x1, [sp, #32]
3464+ ldp x7, x30, [sp], #48
3465+ load_epel_filterh x7, x6
3466+ mov x10, #(MAX_PB_SIZE * 2)
3467+ ld1 {v16.8h, v17.8h, v18.8h, v19.8h}, [sp], x10
3468+ ld1 {v20.8h, v21.8h, v22.8h, v23.8h}, [sp], x10
3469+ ld1 {v24.8h, v25.8h, v26.8h, v27.8h}, [sp], x10
3470+.macro calc src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10, src11, src12, src13, src14, src15
3471+ ld1 {\src12\().8h, \src13\().8h, \src14\().8h, \src15\().8h}, [sp], x10
3472+ calc_epelh v1, \src0, \src4, \src8, \src12
3473+ calc_epelh2 v1, v2, \src0, \src4, \src8, \src12
3474+ calc_epelh v2, \src1, \src5, \src9, \src13
3475+ calc_epelh2 v2, v3, \src1, \src5, \src9, \src13
3476+ calc_epelh v3, \src2, \src6, \src10, \src14
3477+ calc_epelh2 v3, v4, \src2, \src6, \src10, \src14
3478+ calc_epelh v4, \src3, \src7, \src11, \src15
3479+ calc_epelh2 v4, v5, \src3, \src7, \src11, \src15
3480+ ld1 {v5.8h, v6.8h, v7.8h, v8.8h}, [x4], x10
3481+ sqadd v1.8h, v1.8h, v5.8h
3482+ sqadd v2.8h, v2.8h, v6.8h
3483+ sqadd v3.8h, v3.8h, v7.8h
3484+ sqadd v4.8h, v4.8h, v8.8h
3485+ sqrshrun v1.8b, v1.8h, #7
3486+ sqrshrun v2.8b, v2.8h, #7
3487+ sqrshrun v3.8b, v3.8h, #7
3488+ sqrshrun v4.8b, v4.8h, #7
3489+ st1 {v1.8b, v2.8b, v3.8b, v4.8b}, [x0], x1
3490+ subs w5, w5, #1
3491+.endm
3492+1: calc_all16
3493+.purgem calc
3494+2: ldr d8, [sp], #16
3495+ ret
3496+endfunc
3497+
3498+function ff_hevc_put_hevc_epel_bi_hv48_8_neon_i8mm, export=1
3499+ stp x6, x7, [sp, #-80]!
3500+ stp x4, x5, [sp, #16]
3501+ stp x2, x3, [sp, #32]
3502+ stp x0, x1, [sp, #48]
3503+ str x30, [sp, #64]
3504+ bl X(ff_hevc_put_hevc_epel_bi_hv24_8_neon_i8mm)
3505+ ldp x4, x5, [sp, #16]
3506+ ldp x2, x3, [sp, #32]
3507+ ldp x0, x1, [sp, #48]
3508+ ldp x6, x7, [sp], #64
3509+ add x0, x0, #24
3510+ add x2, x2, #24
3511+ add x4, x4, #48
3512+ bl X(ff_hevc_put_hevc_epel_bi_hv24_8_neon_i8mm)
3513+ ldr x30, [sp], #16
3514+ ret
3515+endfunc
3516+
3517+function ff_hevc_put_hevc_epel_bi_hv64_8_neon_i8mm, export=1
3518+ stp x6, x7, [sp, #-80]!
3519+ stp x4, x5, [sp, #16]
3520+ stp x2, x3, [sp, #32]
3521+ stp x0, x1, [sp, #48]
3522+ str x30, [sp, #64]
3523+ bl X(ff_hevc_put_hevc_epel_bi_hv32_8_neon_i8mm)
3524+ ldp x4, x5, [sp, #16]
3525+ ldp x2, x3, [sp, #32]
3526+ ldp x0, x1, [sp, #48]
3527+ ldp x6, x7, [sp], #64
3528+ add x0, x0, #32
3529+ add x2, x2, #32
3530+ add x4, x4, #64
3531+ bl X(ff_hevc_put_hevc_epel_bi_hv32_8_neon_i8mm)
3532+ ldr x30, [sp], #16
3533+ ret
3534+endfunc
3535+
32063536 DISABLE_I8MM
32073537 #endif
32083538
libavcodec/aarch64/hevcdsp_init_aarch64.c+5−0
@@ -168,6 +168,10 @@ NEON8_FNPROTO(epel_bi_v, (uint8_t *dst, ptrdiff_t dststride, · 167 unmodified lines
168168 const uint8_t *src, ptrdiff_t srcstride, const int16_t *src2,
169169 int height, intptr_t mx, intptr_t my, int width),);
170170
171+NEON8_FNPROTO(epel_bi_hv, (uint8_t *dst, ptrdiff_t dststride,
172+ const uint8_t *src, ptrdiff_t srcstride, const int16_t *src2,
173+ int height, intptr_t mx, intptr_t my, int width), _i8mm);
174+
171175 NEON8_FNPROTO(epel_v, (int16_t *dst,
172176 const uint8_t *src, ptrdiff_t srcstride,
173177 int height, intptr_t mx, intptr_t my, int width),);
@@ -354,6 +358,7 @@ av_cold void ff_hevc_dsp_init_aarch64(HEVCDSPContext *c, const int bit_depth) · 180 unmodified lines
354358 NEON8_FNASSIGN(c->put_hevc_epel, 1, 1, epel_hv, _i8mm);
355359 NEON8_FNASSIGN(c->put_hevc_epel_uni, 1, 1, epel_uni_hv, _i8mm);
356360 NEON8_FNASSIGN(c->put_hevc_epel_uni_w, 0, 1, epel_uni_w_h ,_i8mm);
361+ NEON8_FNASSIGN(c->put_hevc_epel_bi, 1, 1, epel_bi_hv, _i8mm);
357362 NEON8_FNASSIGN(c->put_hevc_qpel, 0, 1, qpel_h, _i8mm);
358363 NEON8_FNASSIGN(c->put_hevc_qpel, 1, 1, qpel_hv, _i8mm);
359364 NEON8_FNASSIGN(c->put_hevc_qpel_uni, 1, 1, qpel_uni_hv, _i8mm);

Keyboard shortcuts

General

Show keyboard shortcuts
?
Focus the current search
/

Issues and pull requests

Open filters
f

Issues, pull requests and Inbox

Select next item
j↓
Select previous item
k↑
Open selected item
Enter
Dismiss filters or clear selection
Esc

Repository code

Go to file
t