diff options
Diffstat (limited to '')
| -rw-r--r-- | third_party/dav1d/src/arm/64/itx16.S | 21 |
1 files changed, 15 insertions, 6 deletions
diff --git a/third_party/dav1d/src/arm/64/itx16.S b/third_party/dav1d/src/arm/64/itx16.S index eee3a9636d..31ee9be1b4 100644 --- a/third_party/dav1d/src/arm/64/itx16.S +++ b/third_party/dav1d/src/arm/64/itx16.S @@ -514,13 +514,17 @@ function inv_txfm_add_wht_wht_4x4_16bpc_neon, export=1 b L(itx_4x4_end) endfunc +// HBD inv_txfm_add_4x4_neon deviates from the common pattern with registers +// x0-x4 external parameters +// x5 function pointer to first transform +// x6 function pointer to second transform function inv_txfm_add_4x4_neon movi v30.4s, #0 movi v31.4s, #0 ld1 {v16.4s,v17.4s,v18.4s,v19.4s}, [x2] st1 {v30.4s, v31.4s}, [x2], #32 - blr x4 + blr x5 st1 {v30.4s, v31.4s}, [x2], #32 sqxtn v16.4h, v16.4s @@ -529,7 +533,7 @@ function inv_txfm_add_4x4_neon sqxtn v19.4h, v19.4s transpose_4x4h v16, v17, v18, v19, v20, v21, v22, v23 - blr x5 + blr x6 ld1 {v0.d}[0], [x0], x1 ld1 {v0.d}[1], [x0], x1 @@ -541,7 +545,7 @@ function inv_txfm_add_4x4_neon srshr v18.8h, v18.8h, #4 L(itx_4x4_end): - mvni v31.8h, #0xfc, lsl #8 // 0x3ff + dup v31.8h, w4 sub x0, x0, x1, lsl #2 usqadd v0.8h, v16.8h usqadd v1.8h, v18.8h @@ -579,8 +583,8 @@ function inv_txfm_add_\txfm1\()_\txfm2\()_4x4_16bpc_neon, export=1 b L(itx_4x4_end) 1: .endif - adr x4, inv_\txfm1\()_4s_x4_neon - movrel x5, X(inv_\txfm2\()_4h_x4_neon) + adr x5, inv_\txfm1\()_4s_x4_neon + movrel x6, X(inv_\txfm2\()_4h_x4_neon) b inv_txfm_add_4x4_neon endfunc .endm @@ -1381,6 +1385,10 @@ function inv_txfm_horz\suffix\()_16x4_neon sqrshrn2 v21.8h, v29.4s, #\shift sqrshrn2 v22.8h, v30.4s, #\shift sqrshrn2 v23.8h, v31.4s, #\shift +.if \scale + b L(horz_16x4_epilog) +.else +L(horz_16x4_epilog): transpose_4x8h v16, v17, v18, v19, v4, v5, v6, v7 transpose_4x8h v20, v21, v22, v23, v4, v5, v6, v7 @@ -1389,11 +1397,12 @@ function inv_txfm_horz\suffix\()_16x4_neon .endr ret x14 +.endif endfunc .endm -def_horz_16 scale=0, shift=2 def_horz_16 scale=1, shift=1, suffix=_scale +def_horz_16 scale=0, shift=2 function inv_txfm_add_vert_8x16_neon mov x14, x30 |