[FFmpeg-devel] [PATCH] codec/aarch64/hevc: add transform_luma_neon

[xufuji456]

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 at qq.com>
---
 libavcodec/aarch64/hevcdsp_idct_neon.S    | 50 ++++++++++++++++++++++-
 libavcodec/aarch64/hevcdsp_init_aarch64.c |  2 +
 2 files changed, 51 insertions(+), 1 deletion(-)

diff --git a/libavcodec/aarch64/hevcdsp_idct_neon.S b/libavcodec/aarch64/hevcdsp_idct_neon.S
index 994f0a47b6..504258f7c7 100644
--- a/libavcodec/aarch64/hevcdsp_idct_neon.S
+++ b/libavcodec/aarch64/hevcdsp_idct_neon.S
@@ -889,4 +889,52 @@ idct_dc 16, 8
 idct_dc 16, 10
 
 idct_dc 32, 8
-idct_dc 32, 10
\ No newline at end of file
+idct_dc 32, 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
diff --git a/libavcodec/aarch64/hevcdsp_init_aarch64.c b/libavcodec/aarch64/hevcdsp_init_aarch64.c
index 4cc8732ad3..be1049a2ec 100644
--- a/libavcodec/aarch64/hevcdsp_init_aarch64.c
+++ b/libavcodec/aarch64/hevcdsp_init_aarch64.c
@@ -78,6 +78,7 @@ void ff_hevc_idct_4x4_dc_10_neon(int16_t *coeffs);
 void ff_hevc_idct_8x8_dc_10_neon(int16_t *coeffs);
 void ff_hevc_idct_16x16_dc_10_neon(int16_t *coeffs);
 void ff_hevc_idct_32x32_dc_10_neon(int16_t *coeffs);
+void ff_hevc_transform_luma_4x4_neon_8(int16_t *coeffs);
 void ff_hevc_sao_band_filter_8x8_8_neon(uint8_t *_dst, const uint8_t *_src,
                                   ptrdiff_t stride_dst, ptrdiff_t stride_src,
                                   const int16_t *sao_offset_val, int sao_left_class,
@@ -146,6 +147,7 @@ av_cold void ff_hevc_dsp_init_aarch64(HEVCDSPContext *c, const int bit_depth)
         c->idct_dc[1]                  = ff_hevc_idct_8x8_dc_8_neon;
         c->idct_dc[2]                  = ff_hevc_idct_16x16_dc_8_neon;
         c->idct_dc[3]                  = ff_hevc_idct_32x32_dc_8_neon;
+        c->transform_4x4_luma          = ff_hevc_transform_luma_4x4_neon_8;
         c->sao_band_filter[0]          =
         c->sao_band_filter[1]          =
         c->sao_band_filter[2]          =
-- 
2.32.0 (Apple Git-132)