diff options
| author | J. Dekker <jdek@itanimul.li> | 2022-04-28 14:57:33 +0200 |
|---|---|---|
| committer | J. Dekker <jdek@itanimul.li> | 2022-05-25 08:04:39 +0200 |
| commit | 92f67e40170994dcb7a96ae362d95308f6744294 (patch) | |
| tree | 6e49fba4ddc9e9b014a056676d804662cc7dcc07 /libavcodec/aarch64/hevcdsp_sao_neon.S | |
| parent | d957ee34a6ec998ea00d6d07ac687c5d7a9792a2 (diff) | |
lavc/aarch64: add hevc sao edge 16x16
bench on AWS Graviton:
hevc_sao_edge_16x16_8_c: 1857.0
hevc_sao_edge_16x16_8_neon: 211.0
hevc_sao_edge_32x32_8_c: 7802.2
hevc_sao_edge_32x32_8_neon: 808.2
hevc_sao_edge_48x48_8_c: 16764.2
hevc_sao_edge_48x48_8_neon: 1796.5
hevc_sao_edge_64x64_8_c: 32647.5
hevc_sao_edge_64x64_8_neon: 3118.5
Signed-off-by: J. Dekker <jdek@itanimul.li>
Diffstat (limited to 'libavcodec/aarch64/hevcdsp_sao_neon.S')
| -rw-r--r-- | libavcodec/aarch64/hevcdsp_sao_neon.S | 66 |
1 files changed, 66 insertions, 0 deletions
diff --git a/libavcodec/aarch64/hevcdsp_sao_neon.S b/libavcodec/aarch64/hevcdsp_sao_neon.S index e07e0cea2d..0315c479df 100644 --- a/libavcodec/aarch64/hevcdsp_sao_neon.S +++ b/libavcodec/aarch64/hevcdsp_sao_neon.S @@ -74,3 +74,69 @@ function ff_hevc_sao_band_filter_8x8_8_neon, export=1 bne 1b ret endfunc + +// ASSUMES STRIDE_SRC = 192 +.Lsao_edge_pos: +.word 1 // horizontal +.word 192 // vertical +.word 192 + 1 // 45 degree +.word 192 - 1 // 135 degree + +// ff_hevc_sao_edge_filter_16x16_8_neon(char *dst, char *src, ptrdiff stride_dst, +// int16 *sao_offset_val, int eo, int width, int height) +function ff_hevc_sao_edge_filter_16x16_8_neon, export=1 + adr x7, .Lsao_edge_pos + ld1 {v3.8h}, [x3] // load sao_offset_val + add w5, w5, #0xF + bic w5, w5, #0xF + ldr w4, [x7, w4, uxtw #2] // stride_src + mov v3.h[7], v3.h[0] // reorder to [1,2,0,3,4] + mov v3.h[0], v3.h[1] + mov v3.h[1], v3.h[2] + mov v3.h[2], v3.h[7] + // split 16bit values into two tables + uzp2 v1.16b, v3.16b, v3.16b // sao_offset_val -> upper + uzp1 v0.16b, v3.16b, v3.16b // sao_offset_val -> lower + movi v2.16b, #2 + mov x15, #192 + // strides between end of line and next src/dst + sub x15, x15, x5 // stride_src - width + sub x16, x2, x5 // stride_dst - width + mov x11, x1 // copy base src +1: // new line + mov x14, x5 // copy width + sub x12, x11, x4 // src_a (prev) = src - sao_edge_pos + add x13, x11, x4 // src_b (next) = src + sao_edge_pos +2: // process 16 bytes + ld1 {v3.16b}, [x11], #16 // load src + ld1 {v4.16b}, [x12], #16 // load src_a (prev) + ld1 {v5.16b}, [x13], #16 // load src_b (next) + cmhi v16.16b, v4.16b, v3.16b // (prev > cur) + cmhi v17.16b, v3.16b, v4.16b // (cur > prev) + cmhi v18.16b, v5.16b, v3.16b // (next > cur) + cmhi v19.16b, v3.16b, v5.16b // (cur > next) + sub v20.16b, v16.16b, v17.16b // diff0 = CMP(cur, prev) = (cur > prev) - (cur < prev) + sub v21.16b, v18.16b, v19.16b // diff1 = CMP(cur, next) = (cur > next) - (cur < next) + add v20.16b, v20.16b, v21.16b // diff = diff0 + diff1 + add v20.16b, v20.16b, v2.16b // offset_val = diff + 2 + tbl v16.16b, {v0.16b}, v20.16b + tbl v17.16b, {v1.16b}, v20.16b + uxtl v20.8h, v3.8b // src[0:7] + uxtl2 v21.8h, v3.16b // src[7:15] + zip1 v18.16b, v16.16b, v17.16b // sao_offset_val lower -> + zip2 v19.16b, v16.16b, v17.16b // sao_offset_val upper -> + sqadd v20.8h, v18.8h, v20.8h // + sao_offset_val + sqadd v21.8h, v19.8h, v21.8h + sqxtun v3.8b, v20.8h + sqxtun2 v3.16b, v21.8h + st1 {v3.16b}, [x0], #16 + subs x14, x14, #16 // filtered 16 bytes + b.ne 2b // do we have width to filter? + // no width to filter, setup next line + add x11, x11, x15 // stride src to next line + add x0, x0, x16 // stride dst to next line + subs w6, w6, #1 // filtered line + b.ne 1b // do we have lines to process? + // no lines to filter + ret +endfunc |