[FFmpeg-devel] [PATCH 13/14] swscale/aarch64/range_convert: update neon range_convert functions to new API

[Ramiro Polla]

                                  A55              A76
chrRangeFromJpeg8_1920_c:     28842.4           6346.5
chrRangeFromJpeg8_1920_neon:   5310.9 ( 5.43x)  2264.2 ( 2.80x)
chrRangeToJpeg8_1920_c:       36520.7           9514.0
chrRangeToJpeg8_1920_neon:     6033.2 ( 6.05x)  2645.5 ( 3.60x)
lumRangeFromJpeg8_1920_c:     15387.2           4444.5
lumRangeFromJpeg8_1920_neon:   3148.9 ( 4.89x)  1108.0 ( 4.01x)
lumRangeToJpeg8_1920_c:       19226.4           6015.5
lumRangeToJpeg8_1920_neon:     3866.7 ( 4.97x)  1344.8 ( 4.47x)
---
 libswscale/aarch64/range_convert_neon.S | 63 +++++++++++++------------
 libswscale/aarch64/swscale.c            | 14 +++---
 2 files changed, 41 insertions(+), 36 deletions(-)

diff --git a/libswscale/aarch64/range_convert_neon.S b/libswscale/aarch64/range_convert_neon.S
index 30991ab2a6..3454ee4932 100644
--- a/libswscale/aarch64/range_convert_neon.S
+++ b/libswscale/aarch64/range_convert_neon.S
@@ -20,20 +20,21 @@
 
 #include "libavutil/aarch64/asm.S"
 
-.macro lumConvertRange name, max, mult, offset, shift
-function ff_\name, export=1
-.if \max != 0
-        mov             w3, #\max
-        dup             v24.8h, w3
+.macro lumConvertRange fromto
+function ff_lumRange\fromto\()Jpeg_neon, export=1
+// x0  int16_t *dst
+// w1  int width
+// w2  int amax
+// w3  int coeff
+// x4  int64_t offset
+.ifc \fromto, To
+        dup             v24.8h, w2
 .endif
-        mov             w3, #\mult
         dup             v25.4s, w3
-        movz            w3, #(\offset & 0xffff)
-        movk            w3, #((\offset >> 16) & 0xffff), lsl #16
-        dup             v26.4s, w3
+        dup             v26.4s, w4
 1:
         ld1             {v0.8h}, [x0]
-.if \max != 0
+.ifc \fromto, To
         smin            v0.8h, v0.8h, v24.8h
 .endif
         mov             v16.16b, v26.16b
@@ -42,8 +43,8 @@ function ff_\name, export=1
         sxtl2           v22.4s, v0.8h
         mla             v16.4s, v20.4s, v25.4s
         mla             v18.4s, v22.4s, v25.4s
-        shrn            v0.4h, v16.4s, #\shift
-        shrn2           v0.8h, v18.4s, #\shift
+        shrn            v0.4h, v16.4s, 14
+        shrn2           v0.8h, v18.4s, 14
         subs            w1, w1, #8
         st1             {v0.8h}, [x0], #16
         b.gt            1b
@@ -51,21 +52,23 @@ function ff_\name, export=1
 endfunc
 .endm
 
-.macro chrConvertRange name, max, mult, offset, shift
-function ff_\name, export=1
-.if \max != 0
-        mov             w3, #\max
+.macro chrConvertRange fromto
+function ff_chrRange\fromto\()Jpeg_neon, export=1
+// x0  int16_t *dstU
+// x1  int16_t *dstV
+// w2  int width
+// w3  int amax
+// w4  int coeff
+// x5  int64_t offset
+.ifc \fromto, To
         dup             v24.8h, w3
 .endif
-        mov             w3, #\mult
-        dup             v25.4s, w3
-        movz            w3, #(\offset & 0xffff)
-        movk            w3, #((\offset >> 16) & 0xffff), lsl #16
-        dup             v26.4s, w3
+        dup             v25.4s, w4
+        dup             v26.4s, w5
 1:
         ld1             {v0.8h}, [x0]
         ld1             {v1.8h}, [x1]
-.if \max != 0
+.ifc \fromto, To
         smin            v0.8h, v0.8h, v24.8h
         smin            v1.8h, v1.8h, v24.8h
 .endif
@@ -81,10 +84,10 @@ function ff_\name, export=1
         mla             v17.4s, v21.4s, v25.4s
         mla             v18.4s, v22.4s, v25.4s
         mla             v19.4s, v23.4s, v25.4s
-        shrn            v0.4h, v16.4s, #\shift
-        shrn            v1.4h, v17.4s, #\shift
-        shrn2           v0.8h, v18.4s, #\shift
-        shrn2           v1.8h, v19.4s, #\shift
+        shrn            v0.4h, v16.4s, 14
+        shrn            v1.4h, v17.4s, 14
+        shrn2           v0.8h, v18.4s, 14
+        shrn2           v1.8h, v19.4s, 14
         subs            w2, w2, #8
         st1             {v0.8h}, [x0], #16
         st1             {v1.8h}, [x1], #16
@@ -93,7 +96,7 @@ function ff_\name, export=1
 endfunc
 .endm
 
-lumConvertRange lumRangeToJpeg_neon,   30189, 19077, -39057361, 14
-chrConvertRange chrRangeToJpeg_neon,   30775,  4663,  -9289992, 12
-lumConvertRange lumRangeFromJpeg_neon,     0, 14071,  33561947, 14
-chrConvertRange chrRangeFromJpeg_neon,     0,  1799,   4081085, 11
+lumConvertRange To
+chrConvertRange To
+lumConvertRange From
+chrConvertRange From
diff --git a/libswscale/aarch64/swscale.c b/libswscale/aarch64/swscale.c
index 21788cad5d..55fb81c1e3 100644
--- a/libswscale/aarch64/swscale.c
+++ b/libswscale/aarch64/swscale.c
@@ -218,14 +218,17 @@ NEON_INPUT(bgra32);
 NEON_INPUT(rgb24);
 NEON_INPUT(rgba32);
 
-void ff_lumRangeFromJpeg_neon(int16_t *dst, int width);
-void ff_chrRangeFromJpeg_neon(int16_t *dstU, int16_t *dstV, int width);
-void ff_lumRangeToJpeg_neon(int16_t *dst, int width);
-void ff_chrRangeToJpeg_neon(int16_t *dstU, int16_t *dstV, int width);
+void ff_lumRangeFromJpeg_neon(int16_t *dst, int width,
+                              int amax, int coeff, int64_t offset);
+void ff_chrRangeFromJpeg_neon(int16_t *dstU, int16_t *dstV, int width,
+                              int amax, int coeff, int64_t offset);
+void ff_lumRangeToJpeg_neon(int16_t *dst, int width,
+                            int amax, int coeff, int64_t offset);
+void ff_chrRangeToJpeg_neon(int16_t *dstU, int16_t *dstV, int width,
+                            int amax, int coeff, int64_t offset);
 
 av_cold void ff_sws_init_range_convert_aarch64(SwsContext *c)
 {
-#if 0
     int cpu_flags = av_get_cpu_flags();
 
     if (have_neon(cpu_flags)) {
@@ -239,7 +242,6 @@ av_cold void ff_sws_init_range_convert_aarch64(SwsContext *c)
             }
         }
     }
-#endif
 }
 
 av_cold void ff_sws_init_swscale_aarch64(SwsContext *c)
-- 
2.30.2