[x264-devel] [PATCH 08/11] aarch64: implement x264_sub8x16_dct_dc_neon

[Janne Grunau]

4 times faster than C.
---
 common/aarch64/dct-a.S | 88 +++++++++++++++++++++++++++++---------------------
 common/aarch64/dct.h   |  1 +
 common/dct.c           |  3 ++
 3 files changed, 55 insertions(+), 37 deletions(-)

diff --git a/common/aarch64/dct-a.S b/common/aarch64/dct-a.S
index e33d067..a0e1c40 100644
--- a/common/aarch64/dct-a.S
+++ b/common/aarch64/dct-a.S
@@ -621,56 +621,70 @@ function x264_add16x16_idct_dc_neon, export=1
     ret
 endfunc
 
+.macro sub4x4x2_dct_dc, dst, t0, t1, t2, t3, t4, t5, t6, t7
+    ld1        {\t0\().8b}, [x1], x3
+    ld1        {\t1\().8b}, [x2], x4
+    ld1        {\t2\().8b}, [x1], x3
+    ld1        {\t3\().8b}, [x2], x4
+    usubl       \t0\().8h,  \t0\().8b,  \t1\().8b
+    ld1        {\t4\().8b}, [x1], x3
+    ld1        {\t5\().8b}, [x2], x4
+    usubl       \t1\().8h,  \t2\().8b,  \t3\().8b
+    ld1        {\t6\().8b}, [x1], x3
+    ld1        {\t7\().8b}, [x2], x4
+    add         \dst\().8h, \t0\().8h,  \t1\().8h
+    usubl       \t2\().8h,  \t4\().8b,  \t5\().8b
+    usubl       \t3\().8h,  \t6\().8b,  \t7\().8b
+    add         \dst\().8h, \dst\().8h, \t2\().8h
+    add         \dst\().8h, \dst\().8h, \t3\().8h
+.endm
+
 function x264_sub8x8_dct_dc_neon, export=1
     mov             x3,  #FENC_STRIDE
     mov             x4,  #FDEC_STRIDE
-    ld1        {v16.8b}, [x1], x3
-    ld1        {v17.8b}, [x2], x4
-    usubl       v16.8h,  v16.8b, v17.8b
-    ld1        {v18.8b}, [x1], x3
-    ld1        {v19.8b}, [x2], x4
-    usubl       v17.8h,  v18.8b, v19.8b
-    ld1        {v20.8b}, [x1], x3
-    ld1        {v21.8b}, [x2], x4
-    usubl       v18.8h, v20.8b, v21.8b
-    ld1        {v22.8b}, [x1], x3
-    add         v0.8h,  v16.8h, v17.8h
-    ld1        {v23.8b}, [x2], x4
-    usubl       v19.8h, v22.8b, v23.8b
-    ld1        {v24.8b}, [x1], x3
-    add         v0.8h,  v0.8h,  v18.8h
-    ld1        {v25.8b}, [x2], x4
-    usubl       v20.8h, v24.8b, v25.8b
-    ld1        {v26.8b}, [x1], x3
-    add         v0.8h,  v0.8h,  v19.8h
-    ld1        {v27.8b}, [x2], x4
-    usubl       v21.8h, v26.8b, v27.8b
-    ld1        {v28.8b}, [x1], x3
-    ld1        {v29.8b}, [x2], x4
-    usubl       v22.8h, v28.8b, v29.8b
-    ld1        {v30.8b}, [x1], x3
-    add         v1.8h,  v20.8h, v21.8h
-    ld1        {v31.8b}, [x2], x4
-    usubl       v23.8h, v30.8b, v31.8b
-    add         v1.8h,  v1.8h,  v22.8h
-    add         v1.8h,  v1.8h,  v23.8h
 
+    sub4x4x2_dct_dc  v0, v16, v17, v18, v19, v20, v21, v22, v23
+    sub4x4x2_dct_dc  v1, v24, v25, v26, v27, v28, v29, v30, v31
+
+    transpose   v2.2d,  v3.2d,  v0.2d,  v1.2d
+    SUMSUB_AB   v0.8h,  v1.8h,  v2.8h,  v3.8h
+    transpose   v2.2d,  v3.2d,  v0.2d,  v1.2d
+    SUMSUB_AB   v0.8h,  v1.8h,  v2.8h,  v3.8h
     transpose   v2.2d,  v3.2d,  v0.2d,  v1.2d
 
-    add         v0.8h,  v2.8h,  v3.8h
-    sub         v1.8h,  v2.8h,  v3.8h
+    addp        v0.8h,  v2.8h,  v3.8h
+    addp        v0.8h,  v0.8h,  v0.8h
 
-    transpose   v2.2d,  v3.2d,  v0.2d,  v1.2d
+    st1        {v0.4h}, [x0]
+    ret
+endfunc
+
+function x264_sub8x16_dct_dc_neon, export=1
+    mov             x3,  #FENC_STRIDE
+    mov             x4,  #FDEC_STRIDE
+    sub4x4x2_dct_dc  v0, v16, v17, v18, v19, v20, v21, v22, v23
+    sub4x4x2_dct_dc  v1, v24, v25, v26, v27, v28, v29, v30, v31
+    sub4x4x2_dct_dc  v2, v16, v17, v18, v19, v20, v21, v22, v23
+    sub4x4x2_dct_dc  v3, v24, v25, v26, v27, v28, v29, v30, v31
+
+    addp             v4.8h,  v0.8h,  v2.8h
+    addp             v5.8h,  v1.8h,  v3.8h
 
-    add         v0.8h,  v2.8h,  v3.8h
-    sub         v1.8h,  v2.8h,  v3.8h
+    transpose   v2.4s,  v3.4s,  v4.4s,  v5.4s
+    SUMSUB_AB   v0.8h,  v1.8h,  v2.8h,  v3.8h
+
+    transpose   v2.4s,  v3.4s,  v0.4s,  v1.4s
+    SUMSUB_AB   v0.8h,  v1.8h,  v2.8h,  v3.8h
 
     transpose   v2.2d,  v3.2d,  v0.2d,  v1.2d
+    SUMSUB_AB   v0.8h,  v1.8h,  v2.8h,  v3.8h
+
+    trn1        v2.2d,  v0.2d,  v1.2d
+    trn2        v3.2d,  v1.2d,  v0.2d
 
     addp        v0.8h,  v2.8h,  v3.8h
-    addp        v0.8h,  v0.8h,  v0.8h
 
-    st1        {v0.4h}, [x0]
+    st1        {v0.8h}, [x0]
     ret
 endfunc
 
diff --git a/common/aarch64/dct.h b/common/aarch64/dct.h
index dfd4ec7..c45892f 100644
--- a/common/aarch64/dct.h
+++ b/common/aarch64/dct.h
@@ -40,6 +40,7 @@ void x264_add16x16_idct_neon( uint8_t *p_dst, int16_t dct[16][16] );
 void x264_add8x8_idct_dc_neon( uint8_t *p_dst, int16_t dct[4] );
 void x264_add16x16_idct_dc_neon( uint8_t *p_dst, int16_t dct[16] );
 void x264_sub8x8_dct_dc_neon( int16_t dct[4], uint8_t *pix1, uint8_t *pix2 );
+void x264_sub8x16_dct_dc_neon( int16_t dct[8], uint8_t *pix1, uint8_t *pix2 );
 
 void x264_sub8x8_dct8_neon( int16_t dct[64], uint8_t *pix1, uint8_t *pix2 );
 void x264_sub16x16_dct8_neon( int16_t dct[4][64], uint8_t *pix1, uint8_t *pix2 );
diff --git a/common/dct.c b/common/dct.c
index 8cd1528..ae1b979 100644
--- a/common/dct.c
+++ b/common/dct.c
@@ -747,6 +747,9 @@ void x264_dct_init( int cpu, x264_dct_function_t *dctf )
 
         dctf->add8x8_idct8  = x264_add8x8_idct8_neon;
         dctf->add16x16_idct8= x264_add16x16_idct8_neon;
+#if ARCH_AARCH64
+        dctf->sub8x16_dct_dc= x264_sub8x16_dct_dc_neon;
+#endif
     }
 #endif
 #endif // HIGH_BIT_DEPTH
-- 
2.0.4