History

lavc/aarch64: new optimization for 8-bit hevc_epel_uni_hv

checkasm bench: put_hevc_epel_uni_hv4_8_c: 204.7 put_hevc_epel_uni_hv4_8_i8mm: 70.2 put_hevc_epel_uni_hv6_8_c: 378.2 put_hevc_epel_uni_hv6_8_i8mm: 131.9 put_hevc_epel_uni_hv8_8_c: 637.7 put_hevc_epel_uni_hv8_8_i8mm: 137.9 put_hevc_epel_uni_hv12_8_c: 1301.9 put_hevc_epel_uni_hv12_8_i8mm: 314.2 put_hevc_epel_uni_hv16_8_c: 2203.4 put_hevc_epel_uni_hv16_8_i8mm: 454.7 put_hevc_epel_uni_hv24_8_c: 4848.2 put_hevc_epel_uni_hv24_8_i8mm: 1065.2 put_hevc_epel_uni_hv32_8_c: 8517.4 put_hevc_epel_uni_hv32_8_i8mm: 1898.4 put_hevc_epel_uni_hv48_8_c: 19591.7 put_hevc_epel_uni_hv48_8_i8mm: 4107.2 put_hevc_epel_uni_hv64_8_c: 33880.2 put_hevc_epel_uni_hv64_8_i8mm: 6568.7

Logan Lyu committed 3 years ago

Commit b7a31501 parent c0374f72 files · +299 −0Browse files

Showing the changes to libavcodec (2 of 2 files). Show all

Unified · Split

2 files
  1. libavcodec/aarch64/hevcdsp_epel_neon.S +294 −0
  2. libavcodec/aarch64/hevcdsp_init_aarch64.c +5 −0
libavcodec/aarch64/hevcdsp_epel_neon.S+294−0
@@ -682,6 +682,300 @@ function ff_hevc_put_hevc_epel_h64_8_neon_i8mm, export=1 · 681 unmodified lines
682682 ret
683683 endfunc
684684
685+function ff_hevc_put_hevc_epel_uni_hv4_8_neon_i8mm, export=1
686+ add w10, w4, #3
687+ lsl x10, x10, #7
688+ sub sp, sp, x10 // tmp_array
689+ str x30, [sp, #-48]!
690+ stp x4, x6, [sp, #16]
691+ stp x0, x1, [sp, #32]
692+ add x0, sp, #48
693+ sub x1, x2, x3
694+ mov x2, x3
695+ add w3, w4, #3
696+ mov x4, x5
697+ bl X(ff_hevc_put_hevc_epel_h4_8_neon_i8mm)
698+ ldp x4, x6, [sp, #16]
699+ ldp x0, x1, [sp, #32]
700+ ldr x30, [sp], #48
701+ load_epel_filterh x6, x5
702+ mov x10, #(MAX_PB_SIZE * 2)
703+ ld1 {v16.4h}, [sp], x10
704+ ld1 {v17.4h}, [sp], x10
705+ ld1 {v18.4h}, [sp], x10
706+.macro calc src0, src1, src2, src3
707+ ld1 {\src3\().4h}, [sp], x10
708+ calc_epelh v4, \src0, \src1, \src2, \src3
709+ sqrshrun v4.8b, v4.8h, #6
710+ subs w4, w4, #1
711+ st1 {v4.s}[0], [x0], x1
712+.endm
713+1: calc_all4
714+.purgem calc
715+2: ret
716+endfunc
717+
718+function ff_hevc_put_hevc_epel_uni_hv6_8_neon_i8mm, export=1
719+ add w10, w4, #3
720+ lsl x10, x10, #7
721+ sub sp, sp, x10 // tmp_array
722+ str x30, [sp, #-48]!
723+ stp x4, x6, [sp, #16]
724+ stp x0, x1, [sp, #32]
725+ add x0, sp, #48
726+ sub x1, x2, x3
727+ mov x2, x3
728+ add w3, w4, #3
729+ mov x4, x5
730+ bl X(ff_hevc_put_hevc_epel_h6_8_neon_i8mm)
731+ ldp x4, x6, [sp, #16]
732+ ldp x0, x1, [sp, #32]
733+ ldr x30, [sp], #48
734+ load_epel_filterh x6, x5
735+ sub x1, x1, #4
736+ mov x10, #(MAX_PB_SIZE * 2)
737+ ld1 {v16.8h}, [sp], x10
738+ ld1 {v17.8h}, [sp], x10
739+ ld1 {v18.8h}, [sp], x10
740+.macro calc src0, src1, src2, src3
741+ ld1 {\src3\().8h}, [sp], x10
742+ calc_epelh v4, \src0, \src1, \src2, \src3
743+ calc_epelh2 v4, v5, \src0, \src1, \src2, \src3
744+ sqrshrun v4.8b, v4.8h, #6
745+ st1 {v4.s}[0], [x0], #4
746+ subs w4, w4, #1
747+ st1 {v4.h}[2], [x0], x1
748+.endm
749+1: calc_all4
750+.purgem calc
751+2: ret
752+endfunc
753+
754+function ff_hevc_put_hevc_epel_uni_hv8_8_neon_i8mm, export=1
755+ add w10, w4, #3
756+ lsl x10, x10, #7
757+ sub sp, sp, x10 // tmp_array
758+ str x30, [sp, #-48]!
759+ stp x4, x6, [sp, #16]
760+ stp x0, x1, [sp, #32]
761+ add x0, sp, #48
762+ sub x1, x2, x3
763+ mov x2, x3
764+ add w3, w4, #3
765+ mov x4, x5
766+ bl X(ff_hevc_put_hevc_epel_h8_8_neon_i8mm)
767+ ldp x4, x6, [sp, #16]
768+ ldp x0, x1, [sp, #32]
769+ ldr x30, [sp], #48
770+ load_epel_filterh x6, x5
771+ mov x10, #(MAX_PB_SIZE * 2)
772+ ld1 {v16.8h}, [sp], x10
773+ ld1 {v17.8h}, [sp], x10
774+ ld1 {v18.8h}, [sp], x10
775+.macro calc src0, src1, src2, src3
776+ ld1 {\src3\().8h}, [sp], x10
777+ calc_epelh v4, \src0, \src1, \src2, \src3
778+ calc_epelh2 v4, v5, \src0, \src1, \src2, \src3
779+ sqrshrun v4.8b, v4.8h, #6
780+ subs w4, w4, #1
781+ st1 {v4.8b}, [x0], x1
782+.endm
783+1: calc_all4
784+.purgem calc
785+2: ret
786+endfunc
787+
788+function ff_hevc_put_hevc_epel_uni_hv12_8_neon_i8mm, export=1
789+ add w10, w4, #3
790+ lsl x10, x10, #7
791+ sub sp, sp, x10 // tmp_array
792+ str x30, [sp, #-48]!
793+ stp x4, x6, [sp, #16]
794+ stp x0, x1, [sp, #32]
795+ add x0, sp, #48
796+ sub x1, x2, x3
797+ mov x2, x3
798+ add w3, w4, #3
799+ mov x4, x5
800+ bl X(ff_hevc_put_hevc_epel_h12_8_neon_i8mm)
801+ ldp x4, x6, [sp, #16]
802+ ldp x0, x1, [sp, #32]
803+ ldr x30, [sp], #48
804+ load_epel_filterh x6, x5
805+ sub x1, x1, #8
806+ mov x10, #(MAX_PB_SIZE * 2)
807+ ld1 {v16.8h, v17.8h}, [sp], x10
808+ ld1 {v18.8h, v19.8h}, [sp], x10
809+ ld1 {v20.8h, v21.8h}, [sp], x10
810+.macro calc src0, src1, src2, src3, src4, src5, src6, src7
811+ ld1 {\src6\().8h, \src7\().8h}, [sp], x10
812+ calc_epelh v4, \src0, \src2, \src4, \src6
813+ calc_epelh2 v4, v5, \src0, \src2, \src4, \src6
814+ calc_epelh v5, \src1, \src3, \src5, \src7
815+ sqrshrun v4.8b, v4.8h, #6
816+ sqrshrun2 v4.16b, v5.8h, #6
817+ st1 {v4.8b}, [x0], #8
818+ st1 {v4.s}[2], [x0], x1
819+ subs w4, w4, #1
820+.endm
821+1: calc_all8
822+.purgem calc
823+2: ret
824+endfunc
825+
826+function ff_hevc_put_hevc_epel_uni_hv16_8_neon_i8mm, export=1
827+ add w10, w4, #3
828+ lsl x10, x10, #7
829+ sub sp, sp, x10 // tmp_array
830+ str x30, [sp, #-48]!
831+ stp x4, x6, [sp, #16]
832+ stp x0, x1, [sp, #32]
833+ add x0, sp, #48
834+ sub x1, x2, x3
835+ mov x2, x3
836+ add w3, w4, #3
837+ mov x4, x5
838+ bl X(ff_hevc_put_hevc_epel_h16_8_neon_i8mm)
839+ ldp x4, x6, [sp, #16]
840+ ldp x0, x1, [sp, #32]
841+ ldr x30, [sp], #48
842+ load_epel_filterh x6, x5
843+ mov x10, #(MAX_PB_SIZE * 2)
844+ ld1 {v16.8h, v17.8h}, [sp], x10
845+ ld1 {v18.8h, v19.8h}, [sp], x10
846+ ld1 {v20.8h, v21.8h}, [sp], x10
847+.macro calc src0, src1, src2, src3, src4, src5, src6, src7
848+ ld1 {\src6\().8h, \src7\().8h}, [sp], x10
849+ calc_epelh v4, \src0, \src2, \src4, \src6
850+ calc_epelh2 v4, v5, \src0, \src2, \src4, \src6
851+ calc_epelh v5, \src1, \src3, \src5, \src7
852+ calc_epelh2 v5, v6, \src1, \src3, \src5, \src7
853+ sqrshrun v4.8b, v4.8h, #6
854+ sqrshrun2 v4.16b, v5.8h, #6
855+ subs w4, w4, #1
856+ st1 {v4.16b}, [x0], x1
857+.endm
858+1: calc_all8
859+.purgem calc
860+2: ret
861+endfunc
862+
863+function ff_hevc_put_hevc_epel_uni_hv24_8_neon_i8mm, export=1
864+ add w10, w4, #3
865+ lsl x10, x10, #7
866+ sub sp, sp, x10 // tmp_array
867+ str x30, [sp, #-48]!
868+ stp x4, x6, [sp, #16]
869+ stp x0, x1, [sp, #32]
870+ add x0, sp, #48
871+ sub x1, x2, x3
872+ mov x2, x3
873+ add w3, w4, #3
874+ mov x4, x5
875+ bl X(ff_hevc_put_hevc_epel_h24_8_neon_i8mm)
876+ ldp x4, x6, [sp, #16]
877+ ldp x0, x1, [sp, #32]
878+ ldr x30, [sp], #48
879+ load_epel_filterh x6, x5
880+ mov x10, #(MAX_PB_SIZE * 2)
881+ ld1 {v16.8h, v17.8h, v18.8h}, [sp], x10
882+ ld1 {v19.8h, v20.8h, v21.8h}, [sp], x10
883+ ld1 {v22.8h, v23.8h, v24.8h}, [sp], x10
884+.macro calc src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10, src11
885+ ld1 {\src9\().8h, \src10\().8h, \src11\().8h}, [sp], x10
886+ calc_epelh v4, \src0, \src3, \src6, \src9
887+ calc_epelh2 v4, v5, \src0, \src3, \src6, \src9
888+ calc_epelh v5, \src1, \src4, \src7, \src10
889+ calc_epelh2 v5, v6, \src1, \src4, \src7, \src10
890+ calc_epelh v6, \src2, \src5, \src8, \src11
891+ calc_epelh2 v6, v7, \src2, \src5, \src8, \src11
892+ sqrshrun v4.8b, v4.8h, #6
893+ sqrshrun v5.8b, v5.8h, #6
894+ sqrshrun v6.8b, v6.8h, #6
895+ subs w4, w4, #1
896+ st1 {v4.8b, v5.8b, v6.8b}, [x0], x1
897+.endm
898+1: calc_all12
899+.purgem calc
900+2: ret
901+endfunc
902+
903+function ff_hevc_put_hevc_epel_uni_hv32_8_neon_i8mm, export=1
904+ stp x5, x6, [sp, #-64]!
905+ stp x3, x4, [sp, #16]
906+ stp x1, x2, [sp, #32]
907+ stp x0, x30, [sp, #48]
908+ mov x7, #16
909+ bl X(ff_hevc_put_hevc_epel_uni_hv16_8_neon_i8mm)
910+ ldp x5, x6, [sp]
911+ ldp x3, x4, [sp, #16]
912+ ldp x1, x2, [sp, #32]
913+ ldr x0, [sp, #48]
914+ add x0, x0, #16
915+ add x2, x2, #16
916+ mov x7, #16
917+ bl X(ff_hevc_put_hevc_epel_uni_hv16_8_neon_i8mm)
918+ ldr x30, [sp, #56]
919+ add sp, sp, #64
920+ ret
921+endfunc
922+
923+function ff_hevc_put_hevc_epel_uni_hv48_8_neon_i8mm, export=1
924+ stp x5, x6, [sp, #-64]!
925+ stp x3, x4, [sp, #16]
926+ stp x1, x2, [sp, #32]
927+ stp x0, x30, [sp, #48]
928+ mov x7, #24
929+ bl X(ff_hevc_put_hevc_epel_uni_hv24_8_neon_i8mm)
930+ ldp x5, x6, [sp]
931+ ldp x3, x4, [sp, #16]
932+ ldp x1, x2, [sp, #32]
933+ ldr x0, [sp, #48]
934+ add x0, x0, #24
935+ add x2, x2, #24
936+ mov x7, #24
937+ bl X(ff_hevc_put_hevc_epel_uni_hv24_8_neon_i8mm)
938+ ldr x30, [sp, #56]
939+ add sp, sp, #64
940+ ret
941+endfunc
942+
943+function ff_hevc_put_hevc_epel_uni_hv64_8_neon_i8mm, export=1
944+ stp x5, x6, [sp, #-64]!
945+ stp x3, x4, [sp, #16]
946+ stp x1, x2, [sp, #32]
947+ stp x0, x30, [sp, #48]
948+ mov x7, #16
949+ bl X(ff_hevc_put_hevc_epel_uni_hv16_8_neon_i8mm)
950+ ldp x5, x6, [sp]
951+ ldp x3, x4, [sp, #16]
952+ ldp x1, x2, [sp, #32]
953+ ldr x0, [sp, #48]
954+ add x0, x0, #16
955+ add x2, x2, #16
956+ mov x7, #16
957+ bl X(ff_hevc_put_hevc_epel_uni_hv16_8_neon_i8mm)
958+ ldp x5, x6, [sp]
959+ ldp x3, x4, [sp, #16]
960+ ldp x1, x2, [sp, #32]
961+ ldr x0, [sp, #48]
962+ add x0, x0, #32
963+ add x2, x2, #32
964+ mov x7, #16
965+ bl X(ff_hevc_put_hevc_epel_uni_hv16_8_neon_i8mm)
966+ ldp x5, x6, [sp]
967+ ldp x3, x4, [sp, #16]
968+ ldp x1, x2, [sp, #32]
969+ ldr x0, [sp, #48]
970+ add x0, x0, #48
971+ add x2, x2, #48
972+ mov x7, #16
973+ bl X(ff_hevc_put_hevc_epel_uni_hv16_8_neon_i8mm)
974+ ldr x30, [sp, #56]
975+ add sp, sp, #64
976+ ret
977+endfunc
978+
685979 .macro EPEL_UNI_W_H_HEADER
686980 ldr x12, [sp]
687981 sub x2, x2, #1
libavcodec/aarch64/hevcdsp_init_aarch64.c+5−0
@@ -165,6 +165,10 @@ NEON8_FNPROTO(epel_uni_v, (uint8_t *dst, ptrdiff_t dststride, · 164 unmodified lines
165165 const uint8_t *src, ptrdiff_t srcstride,
166166 int height, intptr_t mx, intptr_t my, int width),);
167167
168+NEON8_FNPROTO(epel_uni_hv, (uint8_t *dst, ptrdiff_t _dststride,
169+ const uint8_t *src, ptrdiff_t srcstride,
170+ int height, intptr_t mx, intptr_t my, int width), _i8mm);
171+
168172 NEON8_FNPROTO(epel_uni_w_v, (uint8_t *_dst, ptrdiff_t _dststride,
169173 const uint8_t *_src, ptrdiff_t _srcstride,
170174 int height, int denom, int wx, int ox,
@@ -298,6 +302,7 @@ av_cold void ff_hevc_dsp_init_aarch64(HEVCDSPContext *c, const int bit_depth) · 127 unmodified lines
298302
299303 if (have_i8mm(cpu_flags)) {
300304 NEON8_FNASSIGN(c->put_hevc_epel, 0, 1, epel_h, _i8mm);
305+ NEON8_FNASSIGN(c->put_hevc_epel_uni, 1, 1, epel_uni_hv, _i8mm);
301306 NEON8_FNASSIGN(c->put_hevc_epel_uni_w, 0, 1, epel_uni_w_h ,_i8mm);
302307 NEON8_FNASSIGN(c->put_hevc_qpel, 0, 1, qpel_h, _i8mm);
303308 NEON8_FNASSIGN(c->put_hevc_qpel_uni_w, 0, 1, qpel_uni_w_h, _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