123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916917918919920921922923924925926927928929930931932933934935936937938939940941942943944945946947948949950951952953954955956957958959960961962963964965966967968969970971972973974975976977978979980981982983984985986987988989990991992993994995996997998999100010011002100310041005100610071008100910101011101210131014101510161017101810191020102110221023102410251026102710281029103010311032103310341035103610371038103910401041104210431044104510461047104810491050105110521053105410551056105710581059106010611062106310641065106610671068106910701071107210731074107510761077107810791080108110821083108410851086108710881089109010911092109310941095 |
- #include "libyuv/row.h"
- #include "libyuv/scale.h"
- #include "libyuv/scale_row.h"
- #ifdef __cplusplus
- namespace libyuv {
- extern "C" {
- #endif
- #if !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
- void ScaleRowDown2_NEON(const uint8_t* src_ptr,
- ptrdiff_t src_stride,
- uint8_t* dst,
- int dst_width) {
- (void)src_stride;
- asm volatile(
- "1: \n"
-
- "ld2 {v0.16b,v1.16b}, [%0], #32 \n"
- "subs %w2, %w2, #16 \n"
- "prfm pldl1keep, [%0, 448] \n"
- "st1 {v1.16b}, [%1], #16 \n"
- "b.gt 1b \n"
- : "+r"(src_ptr),
- "+r"(dst),
- "+r"(dst_width)
- :
- : "v0", "v1"
- );
- }
- void ScaleRowDown2Linear_NEON(const uint8_t* src_ptr,
- ptrdiff_t src_stride,
- uint8_t* dst,
- int dst_width) {
- (void)src_stride;
- asm volatile(
- "1: \n"
-
- "ld2 {v0.16b,v1.16b}, [%0], #32 \n"
- "subs %w2, %w2, #16 \n"
- "urhadd v0.16b, v0.16b, v1.16b \n"
- "st1 {v0.16b}, [%1], #16 \n"
- "prfm pldl1keep, [%0, 448] \n"
- "b.gt 1b \n"
- : "+r"(src_ptr),
- "+r"(dst),
- "+r"(dst_width)
- :
- : "v0", "v1"
- );
- }
- void ScaleRowDown2Box_NEON(const uint8_t* src_ptr,
- ptrdiff_t src_stride,
- uint8_t* dst,
- int dst_width) {
- asm volatile(
-
- "add %1, %1, %0 \n"
- "1: \n"
- "ld1 {v0.16b, v1.16b}, [%0], #32 \n"
- "ld1 {v2.16b, v3.16b}, [%1], #32 \n"
- "subs %w3, %w3, #16 \n"
- "uaddlp v0.8h, v0.16b \n"
- "uaddlp v1.8h, v1.16b \n"
- "uadalp v0.8h, v2.16b \n"
- "uadalp v1.8h, v3.16b \n"
- "rshrn v0.8b, v0.8h, #2 \n"
- "rshrn2 v0.16b, v1.8h, #2 \n"
- "st1 {v0.16b}, [%2], #16 \n"
- "prfm pldl1keep, [%0, 448] \n"
- "prfm pldl1keep, [%1, 448] \n"
- "b.gt 1b \n"
- : "+r"(src_ptr),
- "+r"(src_stride),
- "+r"(dst),
- "+r"(dst_width)
- :
- : "v0", "v1", "v2", "v3"
- );
- }
- void ScaleRowDown4_NEON(const uint8_t* src_ptr,
- ptrdiff_t src_stride,
- uint8_t* dst_ptr,
- int dst_width) {
- (void)src_stride;
- asm volatile(
- "1: \n"
- "ld4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"
- "subs %w2, %w2, #8 \n"
- "st1 {v2.8b}, [%1], #8 \n"
- "prfm pldl1keep, [%0, 448] \n"
- "b.gt 1b \n"
- : "+r"(src_ptr),
- "+r"(dst_ptr),
- "+r"(dst_width)
- :
- : "v0", "v1", "v2", "v3", "memory", "cc");
- }
- void ScaleRowDown4Box_NEON(const uint8_t* src_ptr,
- ptrdiff_t src_stride,
- uint8_t* dst_ptr,
- int dst_width) {
- const uint8_t* src_ptr1 = src_ptr + src_stride;
- const uint8_t* src_ptr2 = src_ptr + src_stride * 2;
- const uint8_t* src_ptr3 = src_ptr + src_stride * 3;
- asm volatile(
- "1: \n"
- "ld1 {v0.16b}, [%0], #16 \n"
- "ld1 {v1.16b}, [%2], #16 \n"
- "ld1 {v2.16b}, [%3], #16 \n"
- "ld1 {v3.16b}, [%4], #16 \n"
- "subs %w5, %w5, #4 \n"
- "uaddlp v0.8h, v0.16b \n"
- "uadalp v0.8h, v1.16b \n"
- "uadalp v0.8h, v2.16b \n"
- "uadalp v0.8h, v3.16b \n"
- "addp v0.8h, v0.8h, v0.8h \n"
- "rshrn v0.8b, v0.8h, #4 \n"
- "st1 {v0.s}[0], [%1], #4 \n"
- "prfm pldl1keep, [%0, 448] \n"
- "prfm pldl1keep, [%2, 448] \n"
- "prfm pldl1keep, [%3, 448] \n"
- "prfm pldl1keep, [%4, 448] \n"
- "b.gt 1b \n"
- : "+r"(src_ptr),
- "+r"(dst_ptr),
- "+r"(src_ptr1),
- "+r"(src_ptr2),
- "+r"(src_ptr3),
- "+r"(dst_width)
- :
- : "v0", "v1", "v2", "v3", "memory", "cc");
- }
- void ScaleRowDown34_NEON(const uint8_t* src_ptr,
- ptrdiff_t src_stride,
- uint8_t* dst_ptr,
- int dst_width) {
- (void)src_stride;
- asm volatile(
- "1: \n"
- "ld4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"
- "subs %w2, %w2, #24 \n"
- "orr v2.16b, v3.16b, v3.16b \n"
- "st3 {v0.8b,v1.8b,v2.8b}, [%1], #24 \n"
- "prfm pldl1keep, [%0, 448] \n"
- "b.gt 1b \n"
- : "+r"(src_ptr),
- "+r"(dst_ptr),
- "+r"(dst_width)
- :
- : "v0", "v1", "v2", "v3", "memory", "cc");
- }
- void ScaleRowDown34_0_Box_NEON(const uint8_t* src_ptr,
- ptrdiff_t src_stride,
- uint8_t* dst_ptr,
- int dst_width) {
- asm volatile(
- "movi v20.8b, #3 \n"
- "add %3, %3, %0 \n"
- "1: \n"
- "ld4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"
- "ld4 {v4.8b,v5.8b,v6.8b,v7.8b}, [%3], #32 \n"
- "subs %w2, %w2, #24 \n"
-
-
-
- "ushll v16.8h, v4.8b, #0 \n"
- "ushll v17.8h, v5.8b, #0 \n"
- "ushll v18.8h, v6.8b, #0 \n"
- "ushll v19.8h, v7.8b, #0 \n"
-
- "umlal v16.8h, v0.8b, v20.8b \n"
- "umlal v17.8h, v1.8b, v20.8b \n"
- "umlal v18.8h, v2.8b, v20.8b \n"
- "umlal v19.8h, v3.8b, v20.8b \n"
-
- "uqrshrn v0.8b, v16.8h, #2 \n"
- "uqrshrn v1.8b, v17.8h, #2 \n"
- "uqrshrn v2.8b, v18.8h, #2 \n"
- "uqrshrn v3.8b, v19.8h, #2 \n"
-
- "ushll v16.8h, v1.8b, #0 \n"
- "umlal v16.8h, v0.8b, v20.8b \n"
- "uqrshrn v0.8b, v16.8h, #2 \n"
-
- "urhadd v1.8b, v1.8b, v2.8b \n"
-
- "ushll v16.8h, v2.8b, #0 \n"
- "umlal v16.8h, v3.8b, v20.8b \n"
- "uqrshrn v2.8b, v16.8h, #2 \n"
- "st3 {v0.8b,v1.8b,v2.8b}, [%1], #24 \n"
- "prfm pldl1keep, [%0, 448] \n"
- "prfm pldl1keep, [%3, 448] \n"
- "b.gt 1b \n"
- : "+r"(src_ptr),
- "+r"(dst_ptr),
- "+r"(dst_width),
- "+r"(src_stride)
- :
- : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17", "v18",
- "v19", "v20", "memory", "cc");
- }
- void ScaleRowDown34_1_Box_NEON(const uint8_t* src_ptr,
- ptrdiff_t src_stride,
- uint8_t* dst_ptr,
- int dst_width) {
- asm volatile(
- "movi v20.8b, #3 \n"
- "add %3, %3, %0 \n"
- "1: \n"
- "ld4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"
- "ld4 {v4.8b,v5.8b,v6.8b,v7.8b}, [%3], #32 \n"
- "subs %w2, %w2, #24 \n"
-
- "urhadd v0.8b, v0.8b, v4.8b \n"
- "urhadd v1.8b, v1.8b, v5.8b \n"
- "urhadd v2.8b, v2.8b, v6.8b \n"
- "urhadd v3.8b, v3.8b, v7.8b \n"
-
- "ushll v4.8h, v1.8b, #0 \n"
- "umlal v4.8h, v0.8b, v20.8b \n"
- "uqrshrn v0.8b, v4.8h, #2 \n"
-
- "urhadd v1.8b, v1.8b, v2.8b \n"
-
- "ushll v4.8h, v2.8b, #0 \n"
- "umlal v4.8h, v3.8b, v20.8b \n"
- "uqrshrn v2.8b, v4.8h, #2 \n"
- "st3 {v0.8b,v1.8b,v2.8b}, [%1], #24 \n"
- "prfm pldl1keep, [%0, 448] \n"
- "prfm pldl1keep, [%3, 448] \n"
- "b.gt 1b \n"
- : "+r"(src_ptr),
- "+r"(dst_ptr),
- "+r"(dst_width),
- "+r"(src_stride)
- :
- : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20", "memory", "cc");
- }
- static const uvec8 kShuf38 = {0, 3, 6, 8, 11, 14, 16, 19,
- 22, 24, 27, 30, 0, 0, 0, 0};
- static const uvec8 kShuf38_2 = {0, 16, 32, 2, 18, 33, 4, 20,
- 34, 6, 22, 35, 0, 0, 0, 0};
- static const vec16 kMult38_Div6 = {65536 / 12, 65536 / 12, 65536 / 12,
- 65536 / 12, 65536 / 12, 65536 / 12,
- 65536 / 12, 65536 / 12};
- static const vec16 kMult38_Div9 = {65536 / 18, 65536 / 18, 65536 / 18,
- 65536 / 18, 65536 / 18, 65536 / 18,
- 65536 / 18, 65536 / 18};
- void ScaleRowDown38_NEON(const uint8_t* src_ptr,
- ptrdiff_t src_stride,
- uint8_t* dst_ptr,
- int dst_width) {
- (void)src_stride;
- asm volatile(
- "ld1 {v3.16b}, [%3] \n"
- "1: \n"
- "ld1 {v0.16b,v1.16b}, [%0], #32 \n"
- "subs %w2, %w2, #12 \n"
- "tbl v2.16b, {v0.16b,v1.16b}, v3.16b \n"
- "st1 {v2.8b}, [%1], #8 \n"
- "st1 {v2.s}[2], [%1], #4 \n"
- "prfm pldl1keep, [%0, 448] \n"
- "b.gt 1b \n"
- : "+r"(src_ptr),
- "+r"(dst_ptr),
- "+r"(dst_width)
- : "r"(&kShuf38)
- : "v0", "v1", "v2", "v3", "memory", "cc");
- }
- void OMITFP ScaleRowDown38_3_Box_NEON(const uint8_t* src_ptr,
- ptrdiff_t src_stride,
- uint8_t* dst_ptr,
- int dst_width) {
- const uint8_t* src_ptr1 = src_ptr + src_stride * 2;
- ptrdiff_t tmp_src_stride = src_stride;
- asm volatile(
- "ld1 {v29.8h}, [%5] \n"
- "ld1 {v30.16b}, [%6] \n"
- "ld1 {v31.8h}, [%7] \n"
- "add %2, %2, %0 \n"
- "1: \n"
-
-
-
-
- "ld4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"
- "ld4 {v4.8b,v5.8b,v6.8b,v7.8b}, [%2], #32 \n"
- "ld4 {v16.8b,v17.8b,v18.8b,v19.8b}, [%3], #32 \n"
- "subs %w4, %w4, #12 \n"
-
-
-
-
- "trn1 v20.8b, v0.8b, v1.8b \n"
- "trn2 v21.8b, v0.8b, v1.8b \n"
- "trn1 v22.8b, v4.8b, v5.8b \n"
- "trn2 v23.8b, v4.8b, v5.8b \n"
- "trn1 v24.8b, v16.8b, v17.8b \n"
- "trn2 v25.8b, v16.8b, v17.8b \n"
-
-
- "trn1 v0.8b, v2.8b, v3.8b \n"
- "trn2 v1.8b, v2.8b, v3.8b \n"
- "trn1 v4.8b, v6.8b, v7.8b \n"
- "trn2 v5.8b, v6.8b, v7.8b \n"
- "trn1 v16.8b, v18.8b, v19.8b \n"
- "trn2 v17.8b, v18.8b, v19.8b \n"
-
-
- "uaddlp v20.4h, v20.8b \n"
- "uaddlp v21.4h, v21.8b \n"
- "uaddlp v22.4h, v22.8b \n"
- "uaddlp v23.4h, v23.8b \n"
- "uaddlp v24.4h, v24.8b \n"
- "uaddlp v25.4h, v25.8b \n"
-
- "uaddlp v1.4h, v1.8b \n"
- "uaddlp v5.4h, v5.8b \n"
- "uaddlp v17.4h, v17.8b \n"
-
- "add v20.4h, v20.4h, v22.4h \n"
- "add v21.4h, v21.4h, v23.4h \n"
- "add v20.4h, v20.4h, v24.4h \n"
- "add v21.4h, v21.4h, v25.4h \n"
- "add v2.4h, v1.4h, v5.4h \n"
- "add v2.4h, v2.4h, v17.4h \n"
-
-
-
- "sqrdmulh v2.8h, v2.8h, v29.8h \n"
- "xtn v2.8b, v2.8h \n"
-
-
-
-
-
-
- "ushll v16.8h, v16.8b, #0 \n"
- "uaddl v0.8h, v0.8b, v4.8b \n"
-
- "add v0.8h, v0.8h, v16.8h \n"
-
-
- "trn1 v1.8h, v0.8h, v0.8h \n"
- "trn2 v4.8h, v0.8h, v0.8h \n"
- "xtn v0.4h, v1.4s \n"
- "xtn v4.4h, v4.4s \n"
-
- "add v20.8h, v20.8h, v0.8h \n"
- "add v21.8h, v21.8h, v4.8h \n"
-
-
-
- "sqrdmulh v0.8h, v20.8h, v31.8h \n"
- "sqrdmulh v1.8h, v21.8h, v31.8h \n"
-
- "tbl v3.16b, {v0.16b, v1.16b, v2.16b}, v30.16b \n"
- "st1 {v3.8b}, [%1], #8 \n"
- "st1 {v3.s}[2], [%1], #4 \n"
- "prfm pldl1keep, [%0, 448] \n"
- "prfm pldl1keep, [%2, 448] \n"
- "prfm pldl1keep, [%3, 448] \n"
- "b.gt 1b \n"
- : "+r"(src_ptr),
- "+r"(dst_ptr),
- "+r"(tmp_src_stride),
- "+r"(src_ptr1),
- "+r"(dst_width)
- : "r"(&kMult38_Div6),
- "r"(&kShuf38_2),
- "r"(&kMult38_Div9)
- : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17", "v18",
- "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v29", "v30", "v31",
- "memory", "cc");
- }
- void ScaleRowDown38_2_Box_NEON(const uint8_t* src_ptr,
- ptrdiff_t src_stride,
- uint8_t* dst_ptr,
- int dst_width) {
-
- ptrdiff_t tmp_src_stride = src_stride;
- asm volatile(
- "ld1 {v30.8h}, [%4] \n"
- "ld1 {v31.16b}, [%5] \n"
- "add %2, %2, %0 \n"
- "1: \n"
-
-
-
-
- "ld4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"
- "ld4 {v4.8b,v5.8b,v6.8b,v7.8b}, [%2], #32 \n"
- "subs %w3, %w3, #12 \n"
-
-
-
-
- "trn1 v16.8b, v0.8b, v1.8b \n"
- "trn2 v17.8b, v0.8b, v1.8b \n"
- "trn1 v18.8b, v4.8b, v5.8b \n"
- "trn2 v19.8b, v4.8b, v5.8b \n"
-
-
- "trn1 v0.8b, v2.8b, v3.8b \n"
- "trn2 v1.8b, v2.8b, v3.8b \n"
- "trn1 v4.8b, v6.8b, v7.8b \n"
- "trn2 v5.8b, v6.8b, v7.8b \n"
-
-
- "uaddlp v16.4h, v16.8b \n"
- "uaddlp v17.4h, v17.8b \n"
- "uaddlp v18.4h, v18.8b \n"
- "uaddlp v19.4h, v19.8b \n"
-
- "uaddlp v1.4h, v1.8b \n"
- "uaddlp v5.4h, v5.8b \n"
-
- "add v16.4h, v16.4h, v18.4h \n"
- "add v17.4h, v17.4h, v19.4h \n"
- "add v2.4h, v1.4h, v5.4h \n"
-
- "uqrshrn v2.8b, v2.8h, #2 \n"
-
-
-
-
-
-
-
- "uaddl v0.8h, v0.8b, v4.8b \n"
-
-
- "trn1 v1.8h, v0.8h, v0.8h \n"
- "trn2 v4.8h, v0.8h, v0.8h \n"
- "xtn v0.4h, v1.4s \n"
- "xtn v4.4h, v4.4s \n"
-
- "add v16.8h, v16.8h, v0.8h \n"
- "add v17.8h, v17.8h, v4.8h \n"
-
-
-
- "sqrdmulh v0.8h, v16.8h, v30.8h \n"
- "sqrdmulh v1.8h, v17.8h, v30.8h \n"
-
-
- "tbl v3.16b, {v0.16b, v1.16b, v2.16b}, v31.16b \n"
- "st1 {v3.8b}, [%1], #8 \n"
- "st1 {v3.s}[2], [%1], #4 \n"
- "prfm pldl1keep, [%0, 448] \n"
- "prfm pldl1keep, [%2, 448] \n"
- "b.gt 1b \n"
- : "+r"(src_ptr),
- "+r"(dst_ptr),
- "+r"(tmp_src_stride),
- "+r"(dst_width)
- : "r"(&kMult38_Div6),
- "r"(&kShuf38_2)
- : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17", "v18",
- "v19", "v30", "v31", "memory", "cc");
- }
- void ScaleAddRow_NEON(const uint8_t* src_ptr,
- uint16_t* dst_ptr,
- int src_width) {
- asm volatile(
- "1: \n"
- "ld1 {v1.8h, v2.8h}, [%1] \n"
- "ld1 {v0.16b}, [%0], #16 \n"
- "uaddw2 v2.8h, v2.8h, v0.16b \n"
- "uaddw v1.8h, v1.8h, v0.8b \n"
- "st1 {v1.8h, v2.8h}, [%1], #32 \n"
- "subs %w2, %w2, #16 \n"
- "prfm pldl1keep, [%0, 448] \n"
- "b.gt 1b \n"
- : "+r"(src_ptr),
- "+r"(dst_ptr),
- "+r"(src_width)
- :
- : "memory", "cc", "v0", "v1", "v2"
- );
- }
- #define LOAD2_DATA8_LANE(n) \
- "lsr %5, %3, #16 \n" \
- "add %6, %1, %5 \n" \
- "add %3, %3, %4 \n" \
- "ld2 {v4.b, v5.b}[" #n "], [%6] \n"
- void ScaleFilterCols_NEON(uint8_t* dst_ptr,
- const uint8_t* src_ptr,
- int dst_width,
- int x,
- int dx) {
- int dx_offset[4] = {0, 1, 2, 3};
- int* tmp = dx_offset;
- const uint8_t* src_tmp = src_ptr;
- int64_t x64 = (int64_t)x;
- int64_t dx64 = (int64_t)dx;
- asm volatile (
- "dup v0.4s, %w3 \n"
- "dup v1.4s, %w4 \n"
- "ld1 {v2.4s}, [%5] \n"
- "shl v3.4s, v1.4s, #2 \n"
- "mul v1.4s, v1.4s, v2.4s \n"
-
- "add v1.4s, v1.4s, v0.4s \n"
-
- "add v2.4s, v1.4s, v3.4s \n"
- "shl v0.4s, v3.4s, #1 \n"
- "1: \n"
- LOAD2_DATA8_LANE(0)
- LOAD2_DATA8_LANE(1)
- LOAD2_DATA8_LANE(2)
- LOAD2_DATA8_LANE(3)
- LOAD2_DATA8_LANE(4)
- LOAD2_DATA8_LANE(5)
- LOAD2_DATA8_LANE(6)
- LOAD2_DATA8_LANE(7)
- "mov v6.16b, v1.16b \n"
- "mov v7.16b, v2.16b \n"
- "uzp1 v6.8h, v6.8h, v7.8h \n"
- "ushll v4.8h, v4.8b, #0 \n"
- "ushll v5.8h, v5.8b, #0 \n"
- "ssubl v16.4s, v5.4h, v4.4h \n"
- "ssubl2 v17.4s, v5.8h, v4.8h \n"
- "ushll v7.4s, v6.4h, #0 \n"
- "ushll2 v6.4s, v6.8h, #0 \n"
- "mul v16.4s, v16.4s, v7.4s \n"
- "mul v17.4s, v17.4s, v6.4s \n"
- "rshrn v6.4h, v16.4s, #16 \n"
- "rshrn2 v6.8h, v17.4s, #16 \n"
- "add v4.8h, v4.8h, v6.8h \n"
- "xtn v4.8b, v4.8h \n"
- "st1 {v4.8b}, [%0], #8 \n"
- "add v1.4s, v1.4s, v0.4s \n"
- "add v2.4s, v2.4s, v0.4s \n"
- "subs %w2, %w2, #8 \n"
- "b.gt 1b \n"
- : "+r"(dst_ptr),
- "+r"(src_ptr),
- "+r"(dst_width),
- "+r"(x64),
- "+r"(dx64),
- "+r"(tmp),
- "+r"(src_tmp)
- :
- : "memory", "cc", "v0", "v1", "v2", "v3",
- "v4", "v5", "v6", "v7", "v16", "v17"
- );
- }
- #undef LOAD2_DATA8_LANE
- void ScaleFilterRows_NEON(uint8_t* dst_ptr,
- const uint8_t* src_ptr,
- ptrdiff_t src_stride,
- int dst_width,
- int source_y_fraction) {
- int y_fraction = 256 - source_y_fraction;
- asm volatile(
- "cmp %w4, #0 \n"
- "b.eq 100f \n"
- "add %2, %2, %1 \n"
- "cmp %w4, #64 \n"
- "b.eq 75f \n"
- "cmp %w4, #128 \n"
- "b.eq 50f \n"
- "cmp %w4, #192 \n"
- "b.eq 25f \n"
- "dup v5.8b, %w4 \n"
- "dup v4.8b, %w5 \n"
-
- "1: \n"
- "ld1 {v0.16b}, [%1], #16 \n"
- "ld1 {v1.16b}, [%2], #16 \n"
- "subs %w3, %w3, #16 \n"
- "umull v6.8h, v0.8b, v4.8b \n"
- "umull2 v7.8h, v0.16b, v4.16b \n"
- "umlal v6.8h, v1.8b, v5.8b \n"
- "umlal2 v7.8h, v1.16b, v5.16b \n"
- "rshrn v0.8b, v6.8h, #8 \n"
- "rshrn2 v0.16b, v7.8h, #8 \n"
- "st1 {v0.16b}, [%0], #16 \n"
- "prfm pldl1keep, [%1, 448] \n"
- "prfm pldl1keep, [%2, 448] \n"
- "b.gt 1b \n"
- "b 99f \n"
-
- "25: \n"
- "ld1 {v0.16b}, [%1], #16 \n"
- "ld1 {v1.16b}, [%2], #16 \n"
- "subs %w3, %w3, #16 \n"
- "urhadd v0.16b, v0.16b, v1.16b \n"
- "urhadd v0.16b, v0.16b, v1.16b \n"
- "st1 {v0.16b}, [%0], #16 \n"
- "prfm pldl1keep, [%1, 448] \n"
- "prfm pldl1keep, [%2, 448] \n"
- "b.gt 25b \n"
- "b 99f \n"
-
- "50: \n"
- "ld1 {v0.16b}, [%1], #16 \n"
- "ld1 {v1.16b}, [%2], #16 \n"
- "subs %w3, %w3, #16 \n"
- "urhadd v0.16b, v0.16b, v1.16b \n"
- "st1 {v0.16b}, [%0], #16 \n"
- "prfm pldl1keep, [%1, 448] \n"
- "prfm pldl1keep, [%2, 448] \n"
- "b.gt 50b \n"
- "b 99f \n"
-
- "75: \n"
- "ld1 {v1.16b}, [%1], #16 \n"
- "ld1 {v0.16b}, [%2], #16 \n"
- "subs %w3, %w3, #16 \n"
- "urhadd v0.16b, v0.16b, v1.16b \n"
- "urhadd v0.16b, v0.16b, v1.16b \n"
- "st1 {v0.16b}, [%0], #16 \n"
- "prfm pldl1keep, [%1, 448] \n"
- "prfm pldl1keep, [%2, 448] \n"
- "b.gt 75b \n"
- "b 99f \n"
-
- "100: \n"
- "ld1 {v0.16b}, [%1], #16 \n"
- "subs %w3, %w3, #16 \n"
- "st1 {v0.16b}, [%0], #16 \n"
- "prfm pldl1keep, [%1, 448] \n"
- "b.gt 100b \n"
- "99: \n"
- "st1 {v0.b}[15], [%0] \n"
- : "+r"(dst_ptr),
- "+r"(src_ptr),
- "+r"(src_stride),
- "+r"(dst_width),
- "+r"(source_y_fraction),
- "+r"(y_fraction)
- :
- : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "memory", "cc");
- }
- void ScaleARGBRowDown2_NEON(const uint8_t* src_ptr,
- ptrdiff_t src_stride,
- uint8_t* dst,
- int dst_width) {
- (void)src_stride;
- asm volatile(
- "1: \n"
-
- "ld4 {v0.4s,v1.4s,v2.4s,v3.4s}, [%0], #64 \n"
- "subs %w2, %w2, #8 \n"
- "mov v2.16b, v3.16b \n"
- "st2 {v1.4s,v2.4s}, [%1], #32 \n"
- "prfm pldl1keep, [%0, 448] \n"
- "b.gt 1b \n"
- : "+r"(src_ptr),
- "+r"(dst),
- "+r"(dst_width)
- :
- : "memory", "cc", "v0", "v1", "v2", "v3"
- );
- }
- void ScaleARGBRowDown2Linear_NEON(const uint8_t* src_argb,
- ptrdiff_t src_stride,
- uint8_t* dst_argb,
- int dst_width) {
- (void)src_stride;
- asm volatile(
- "1: \n"
-
- "ld4 {v0.4s,v1.4s,v2.4s,v3.4s}, [%0], #64 \n"
- "subs %w2, %w2, #8 \n"
- "urhadd v0.16b, v0.16b, v1.16b \n"
- "urhadd v1.16b, v2.16b, v3.16b \n"
- "st2 {v0.4s,v1.4s}, [%1], #32 \n"
- "prfm pldl1keep, [%0, 448] \n"
- "b.gt 1b \n"
- : "+r"(src_argb),
- "+r"(dst_argb),
- "+r"(dst_width)
- :
- : "memory", "cc", "v0", "v1", "v2", "v3"
- );
- }
- void ScaleARGBRowDown2Box_NEON(const uint8_t* src_ptr,
- ptrdiff_t src_stride,
- uint8_t* dst,
- int dst_width) {
- asm volatile(
-
- "add %1, %1, %0 \n"
- "1: \n"
- "ld4 {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"
- "subs %w3, %w3, #8 \n"
- "uaddlp v0.8h, v0.16b \n"
- "uaddlp v1.8h, v1.16b \n"
- "uaddlp v2.8h, v2.16b \n"
- "uaddlp v3.8h, v3.16b \n"
- "ld4 {v16.16b,v17.16b,v18.16b,v19.16b}, [%1], #64 \n"
- "uadalp v0.8h, v16.16b \n"
- "uadalp v1.8h, v17.16b \n"
- "uadalp v2.8h, v18.16b \n"
- "uadalp v3.8h, v19.16b \n"
- "rshrn v0.8b, v0.8h, #2 \n"
- "rshrn v1.8b, v1.8h, #2 \n"
- "rshrn v2.8b, v2.8h, #2 \n"
- "rshrn v3.8b, v3.8h, #2 \n"
- "st4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n"
- "prfm pldl1keep, [%0, 448] \n"
- "prfm pldl1keep, [%1, 448] \n"
- "b.gt 1b \n"
- : "+r"(src_ptr),
- "+r"(src_stride),
- "+r"(dst),
- "+r"(dst_width)
- :
- : "memory", "cc", "v0", "v1", "v2", "v3", "v16", "v17", "v18", "v19");
- }
- void ScaleARGBRowDownEven_NEON(const uint8_t* src_argb,
- ptrdiff_t src_stride,
- int src_stepx,
- uint8_t* dst_argb,
- int dst_width) {
- (void)src_stride;
- asm volatile(
- "1: \n"
- "ld1 {v0.s}[0], [%0], %3 \n"
- "ld1 {v0.s}[1], [%0], %3 \n"
- "ld1 {v0.s}[2], [%0], %3 \n"
- "ld1 {v0.s}[3], [%0], %3 \n"
- "subs %w2, %w2, #4 \n"
- "st1 {v0.16b}, [%1], #16 \n"
- "prfm pldl1keep, [%0, 448] \n"
- "b.gt 1b \n"
- : "+r"(src_argb),
- "+r"(dst_argb),
- "+r"(dst_width)
- : "r"((int64_t)(src_stepx * 4))
- : "memory", "cc", "v0");
- }
- void ScaleARGBRowDownEvenBox_NEON(const uint8_t* src_argb,
- ptrdiff_t src_stride,
- int src_stepx,
- uint8_t* dst_argb,
- int dst_width) {
- asm volatile(
- "add %1, %1, %0 \n"
- "1: \n"
- "ld1 {v0.8b}, [%0], %4 \n"
- "ld1 {v1.8b}, [%1], %4 \n"
- "ld1 {v2.8b}, [%0], %4 \n"
- "ld1 {v3.8b}, [%1], %4 \n"
- "ld1 {v4.8b}, [%0], %4 \n"
- "ld1 {v5.8b}, [%1], %4 \n"
- "ld1 {v6.8b}, [%0], %4 \n"
- "ld1 {v7.8b}, [%1], %4 \n"
- "uaddl v0.8h, v0.8b, v1.8b \n"
- "uaddl v2.8h, v2.8b, v3.8b \n"
- "uaddl v4.8h, v4.8b, v5.8b \n"
- "uaddl v6.8h, v6.8b, v7.8b \n"
- "mov v16.d[1], v0.d[1] \n"
- "mov v0.d[1], v2.d[0] \n"
- "mov v2.d[0], v16.d[1] \n"
- "mov v16.d[1], v4.d[1] \n"
- "mov v4.d[1], v6.d[0] \n"
- "mov v6.d[0], v16.d[1] \n"
- "add v0.8h, v0.8h, v2.8h \n"
- "add v4.8h, v4.8h, v6.8h \n"
- "rshrn v0.8b, v0.8h, #2 \n"
- "rshrn2 v0.16b, v4.8h, #2 \n"
- "subs %w3, %w3, #4 \n"
- "st1 {v0.16b}, [%2], #16 \n"
- "prfm pldl1keep, [%0, 448] \n"
- "prfm pldl1keep, [%1, 448] \n"
- "b.gt 1b \n"
- : "+r"(src_argb),
- "+r"(src_stride),
- "+r"(dst_argb),
- "+r"(dst_width)
- : "r"((int64_t)(src_stepx * 4))
- : "memory", "cc", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16");
- }
- #define LOAD1_DATA32_LANE(vn, n) \
- "lsr %5, %3, #16 \n" \
- "add %6, %1, %5, lsl #2 \n" \
- "add %3, %3, %4 \n" \
- "ld1 {" #vn ".s}[" #n "], [%6] \n"
- void ScaleARGBCols_NEON(uint8_t* dst_argb,
- const uint8_t* src_argb,
- int dst_width,
- int x,
- int dx) {
- const uint8_t* src_tmp = src_argb;
- int64_t x64 = (int64_t)x;
- int64_t dx64 = (int64_t)dx;
- int64_t tmp64;
- asm volatile(
- "1: \n"
-
- LOAD1_DATA32_LANE(v0, 0)
- LOAD1_DATA32_LANE(v0, 1)
- LOAD1_DATA32_LANE(v0, 2)
- LOAD1_DATA32_LANE(v0, 3)
- LOAD1_DATA32_LANE(v1, 0)
- LOAD1_DATA32_LANE(v1, 1)
- LOAD1_DATA32_LANE(v1, 2)
- LOAD1_DATA32_LANE(v1, 3)
-
- "st1 {v0.4s, v1.4s}, [%0], #32 \n"
- "subs %w2, %w2, #8 \n"
- "prfm pldl1keep, [%1, 448] \n"
- "b.gt 1b \n"
- : "+r"(dst_argb),
- "+r"(src_argb),
- "+r"(dst_width),
- "+r"(x64),
- "+r"(dx64),
- "=&r"(tmp64),
- "+r"(src_tmp)
- :
- : "memory", "cc", "v0", "v1");
- }
- #undef LOAD1_DATA32_LANE
- #define LOAD2_DATA32_LANE(vn1, vn2, n) \
- "lsr %5, %3, #16 \n" \
- "add %6, %1, %5, lsl #2 \n" \
- "add %3, %3, %4 \n" \
- "ld2 {" #vn1 ".s, " #vn2 ".s}[" #n "], [%6] \n"
- void ScaleARGBFilterCols_NEON(uint8_t* dst_argb,
- const uint8_t* src_argb,
- int dst_width,
- int x,
- int dx) {
- int dx_offset[4] = {0, 1, 2, 3};
- int* tmp = dx_offset;
- const uint8_t* src_tmp = src_argb;
- int64_t x64 = (int64_t)x;
- int64_t dx64 = (int64_t)dx;
- asm volatile (
- "dup v0.4s, %w3 \n"
- "dup v1.4s, %w4 \n"
- "ld1 {v2.4s}, [%5] \n"
- "shl v6.4s, v1.4s, #2 \n"
- "mul v1.4s, v1.4s, v2.4s \n"
- "movi v3.16b, #0x7f \n"
- "movi v4.8h, #0x7f \n"
-
- "add v5.4s, v1.4s, v0.4s \n"
- "1: \n"
-
-
- LOAD2_DATA32_LANE(v0, v1, 0)
- LOAD2_DATA32_LANE(v0, v1, 1)
- LOAD2_DATA32_LANE(v0, v1, 2)
- LOAD2_DATA32_LANE(v0, v1, 3)
- "shrn v2.4h, v5.4s, #9 \n"
- "and v2.8b, v2.8b, v4.8b \n"
- "dup v16.8b, v2.b[0] \n"
- "dup v17.8b, v2.b[2] \n"
- "dup v18.8b, v2.b[4] \n"
- "dup v19.8b, v2.b[6] \n"
- "ext v2.8b, v16.8b, v17.8b, #4 \n"
- "ext v17.8b, v18.8b, v19.8b, #4 \n"
- "ins v2.d[1], v17.d[0] \n"
- "eor v7.16b, v2.16b, v3.16b \n"
- "umull v16.8h, v0.8b, v7.8b \n"
- "umull2 v17.8h, v0.16b, v7.16b \n"
- "umull v18.8h, v1.8b, v2.8b \n"
- "umull2 v19.8h, v1.16b, v2.16b \n"
- "add v16.8h, v16.8h, v18.8h \n"
- "add v17.8h, v17.8h, v19.8h \n"
- "shrn v0.8b, v16.8h, #7 \n"
- "shrn2 v0.16b, v17.8h, #7 \n"
- "st1 {v0.4s}, [%0], #16 \n"
- "add v5.4s, v5.4s, v6.4s \n"
- "subs %w2, %w2, #4 \n"
- "prfm pldl1keep, [%1, 448] \n"
- "b.gt 1b \n"
- : "+r"(dst_argb),
- "+r"(src_argb),
- "+r"(dst_width),
- "+r"(x64),
- "+r"(dx64),
- "+r"(tmp),
- "+r"(src_tmp)
- :
- : "memory", "cc", "v0", "v1", "v2", "v3", "v4", "v5",
- "v6", "v7", "v16", "v17", "v18", "v19"
- );
- }
- #undef LOAD2_DATA32_LANE
- void ScaleRowDown2Box_16_NEON(const uint16_t* src_ptr,
- ptrdiff_t src_stride,
- uint16_t* dst,
- int dst_width) {
- asm volatile(
-
- "add %1, %0, %1, lsl #1 \n"
- "1: \n"
- "ld1 {v0.8h, v1.8h}, [%0], #32 \n"
- "ld1 {v2.8h, v3.8h}, [%1], #32 \n"
- "subs %w3, %w3, #8 \n"
- "uaddlp v0.4s, v0.8h \n"
- "uaddlp v1.4s, v1.8h \n"
- "uadalp v0.4s, v2.8h \n"
- "uadalp v1.4s, v3.8h \n"
- "rshrn v0.4h, v0.4s, #2 \n"
- "rshrn2 v0.8h, v1.4s, #2 \n"
- "st1 {v0.8h}, [%2], #16 \n"
- "prfm pldl1keep, [%0, 448] \n"
- "prfm pldl1keep, [%1, 448] \n"
- "b.gt 1b \n"
- : "+r"(src_ptr),
- "+r"(src_stride),
- "+r"(dst),
- "+r"(dst_width)
- :
- : "v0", "v1", "v2", "v3"
- );
- }
- void ScaleRowUp2_16_NEON(const uint16_t* src_ptr,
- ptrdiff_t src_stride,
- uint16_t* dst,
- int dst_width) {
- asm volatile(
- "add %1, %0, %1, lsl #1 \n"
- "movi v0.8h, #9 \n"
- "movi v1.4s, #3 \n"
- "1: \n"
- "ld1 {v3.8h}, [%0], %4 \n"
- "ld1 {v4.8h}, [%0], %5 \n"
- "ld1 {v5.8h}, [%1], %4 \n"
- "ld1 {v6.8h}, [%1], %5 \n"
- "subs %w3, %w3, #16 \n"
- "umull v16.4s, v3.4h, v0.4h \n"
- "umull2 v7.4s, v3.8h, v0.8h \n"
- "umull v18.4s, v4.4h, v0.4h \n"
- "umull2 v17.4s, v4.8h, v0.8h \n"
- "uaddw v16.4s, v16.4s, v6.4h \n"
- "uaddl2 v19.4s, v6.8h, v3.8h \n"
- "uaddl v3.4s, v6.4h, v3.4h \n"
- "uaddw2 v6.4s, v7.4s, v6.8h \n"
- "uaddl2 v7.4s, v5.8h, v4.8h \n"
- "uaddl v4.4s, v5.4h, v4.4h \n"
- "uaddw v18.4s, v18.4s, v5.4h \n"
- "mla v16.4s, v4.4s, v1.4s \n"
- "mla v18.4s, v3.4s, v1.4s \n"
- "mla v6.4s, v7.4s, v1.4s \n"
- "uaddw2 v4.4s, v17.4s, v5.8h \n"
- "uqrshrn v16.4h, v16.4s, #4 \n"
- "mla v4.4s, v19.4s, v1.4s \n"
- "uqrshrn2 v16.8h, v6.4s, #4 \n"
- "uqrshrn v17.4h, v18.4s, #4 \n"
- "uqrshrn2 v17.8h, v4.4s, #4 \n"
- "st2 {v16.8h-v17.8h}, [%2], #32 \n"
- "prfm pldl1keep, [%0, 448] \n"
- "prfm pldl1keep, [%1, 448] \n"
- "b.gt 1b \n"
- : "+r"(src_ptr),
- "+r"(src_stride),
- "+r"(dst),
- "+r"(dst_width)
- : "r"(2LL),
- "r"(14LL)
- : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17", "v18",
- "v19"
- );
- }
- #endif
- #ifdef __cplusplus
- }
- }
- #endif
|