diff options
| author | xufuji456 <839789740@qq.com> | 2023-04-13 21:34:47 +0800 |
|---|---|---|
| committer | Martin Storsjö <martin@martin.st> | 2023-04-14 12:07:57 +0300 |
| commit | bd2f00f665cc964fc1942518cdf27bd6e8b6d388 (patch) | |
| tree | b5a09d6de8dcb6ca2ede30f17ed0d08df1eed5df /libavcodec/aarch64/hevcdsp_idct_neon.S | |
| parent | 4eaaa38d3dfb8863a62f3646a62e4098b1c078d5 (diff) | |
| download | ffmpeg-bd2f00f665cc964fc1942518cdf27bd6e8b6d388.tar.gz | |
codec/aarch64/hevc: add transform_luma_neon
got 56% speed up (run_count=1000, CPU=Cortex A53)
transform_4x4_luma_neon: 45 transform_4x4_luma_c: 103
Signed-off-by: xufuji456 <839789740@qq.com>
Signed-off-by: Martin Storsjö <martin@martin.st>
Diffstat (limited to 'libavcodec/aarch64/hevcdsp_idct_neon.S')
| -rw-r--r-- | libavcodec/aarch64/hevcdsp_idct_neon.S | 48 |
1 files changed, 48 insertions, 0 deletions
diff --git a/libavcodec/aarch64/hevcdsp_idct_neon.S b/libavcodec/aarch64/hevcdsp_idct_neon.S index 976dec236a..b7f23386a4 100644 --- a/libavcodec/aarch64/hevcdsp_idct_neon.S +++ b/libavcodec/aarch64/hevcdsp_idct_neon.S @@ -842,6 +842,54 @@ tr_32x4 secondpass_10, 20 - 10 idct_32x32 8 idct_32x32 10 +.macro tr4_luma_shift r0, r1, r2, r3, shift + saddl v0.4s, \r0, \r2 // c0 = src0 + src2 + saddl v1.4s, \r2, \r3 // c1 = src2 + src3 + ssubl v2.4s, \r0, \r3 // c2 = src0 - src3 + smull v3.4s, \r1, v21.4h // c3 = 74 * src1 + + saddl v7.4s, \r0, \r3 // src0 + src3 + ssubw v7.4s, v7.4s, \r2 // src0 - src2 + src3 + mul v7.4s, v7.4s, v18.4s // dst2 = 74 * (src0 - src2 + src3) + + mul v5.4s, v0.4s, v19.4s // 29 * c0 + mul v6.4s, v1.4s, v20.4s // 55 * c1 + add v5.4s, v5.4s, v6.4s // 29 * c0 + 55 * c1 + add v5.4s, v5.4s, v3.4s // dst0 = 29 * c0 + 55 * c1 + c3 + + mul v1.4s, v1.4s, v19.4s // 29 * c1 + mul v6.4s, v2.4s, v20.4s // 55 * c2 + sub v6.4s, v6.4s, v1.4s // 55 * c2 - 29 * c1 + add v6.4s, v6.4s, v3.4s // dst1 = 55 * c2 - 29 * c1 + c3 + + mul v0.4s, v0.4s, v20.4s // 55 * c0 + mul v2.4s, v2.4s, v19.4s // 29 * c2 + add v0.4s, v0.4s, v2.4s // 55 * c0 + 29 * c2 + sub v0.4s, v0.4s, v3.4s // dst3 = 55 * c0 + 29 * c2 - c3 + + sqrshrn \r0, v5.4s, \shift + sqrshrn \r1, v6.4s, \shift + sqrshrn \r2, v7.4s, \shift + sqrshrn \r3, v0.4s, \shift +.endm + +function ff_hevc_transform_luma_4x4_neon_8, export=1 + ld1 {v28.4h-v31.4h}, [x0] + movi v18.4s, #74 + movi v19.4s, #29 + movi v20.4s, #55 + movi v21.4h, #74 + + tr4_luma_shift v28.4h, v29.4h, v30.4h, v31.4h, #7 + transpose_4x4H v28, v29, v30, v31, v22, v23, v24, v25 + + tr4_luma_shift v28.4h, v29.4h, v30.4h, v31.4h, #12 + transpose_4x4H v28, v29, v30, v31, v22, v23, v24, v25 + + st1 {v28.4h-v31.4h}, [x0] + ret +endfunc + // void ff_hevc_idct_NxN_dc_DEPTH_neon(int16_t *coeffs) .macro idct_dc size, bitdepth function ff_hevc_idct_\size\()x\size\()_dc_\bitdepth\()_neon, export=1 |