[FFmpeg-devel] [PATCH] swscale/aarch64/rgb24toyv12: skip early right shift by 2

[FFmpeg-devel] [PATCH] swscale/aarch64/rgb24toyv12: skip early right shift by 2

[Krzysztof Pyrkosz via ffmpeg-devel]

It's a minor improvement that shaves off 5-8% from the execution time.
Instead of shifting by 2 right away and by 7 soon after, shift by 9 one
time.

Times before and after:

A78:
rgb24toyv12_16_200_neon:                              5366.8 ( 3.62x)
rgb24toyv12_128_60_neon:                             13574.0 ( 3.34x)
rgb24toyv12_512_16_neon:                             14463.8 ( 3.33x)
rgb24toyv12_1920_4_neon:                             13508.2 ( 3.34x)
rgb24toyv12_1920_4_negstride_neon:                   13525.0 ( 3.34x)

rgb24toyv12_16_200_neon:                              5293.8 ( 3.66x)
rgb24toyv12_128_60_neon:                             12955.0 ( 3.50x)
rgb24toyv12_512_16_neon:                             13784.0 ( 3.50x)
rgb24toyv12_1920_4_neon:                             12900.8 ( 3.49x)
rgb24toyv12_1920_4_negstride_neon:                   12902.8 ( 3.49x)


A72:
rgb24toyv12_16_200_neon:                              9695.8 ( 2.50x)
rgb24toyv12_128_60_neon:                             20286.6 ( 2.70x)
rgb24toyv12_512_16_neon:                             22276.6 ( 2.57x)
rgb24toyv12_1920_4_neon:                             19154.1 ( 2.77x)
rgb24toyv12_1920_4_negstride_neon:                   19055.1 ( 2.78x)

rgb24toyv12_16_200_neon:                              9214.8 ( 2.65x)
rgb24toyv12_128_60_neon:                             20731.5 ( 2.65x)
rgb24toyv12_512_16_neon:                             21145.0 ( 2.70x)
rgb24toyv12_1920_4_neon:                             17586.5 ( 2.99x)
rgb24toyv12_1920_4_negstride_neon:                   17571.0 ( 2.98x)


A53:
rgb24toyv12_16_200_neon:                             12880.4 ( 3.76x)
rgb24toyv12_128_60_neon:                             27776.3 ( 3.94x)
rgb24toyv12_512_16_neon:                             29411.3 ( 3.94x)
rgb24toyv12_1920_4_neon:                             27253.1 ( 3.98x)
rgb24toyv12_1920_4_negstride_neon:                   27474.3 ( 3.95x)

rgb24toyv12_16_200_neon:                             12196.3 ( 3.95x)
rgb24toyv12_128_60_neon:                             26943.1 ( 4.07x)
rgb24toyv12_512_16_neon:                             28642.3 ( 4.07x)
rgb24toyv12_1920_4_neon:                             26676.6 ( 4.08x)
rgb24toyv12_1920_4_negstride_neon:                   26713.8 ( 4.07x)

---
 libswscale/aarch64/rgb2rgb_neon.S | 24 +++++++++++-------------
 1 file changed, 11 insertions(+), 13 deletions(-)

diff --git a/libswscale/aarch64/rgb2rgb_neon.S b/libswscale/aarch64/rgb2rgb_neon.S
index 22ecdf7ac8..97dd781378 100644
--- a/libswscale/aarch64/rgb2rgb_neon.S
+++ b/libswscale/aarch64/rgb2rgb_neon.S
@@ -99,15 +99,16 @@ endconst
 
 // convert rgb to 16-bit y, u, or v
 // uses v3 and v4
-.macro rgbconv16 dst, b, g, r, bc, gc, rc
+
+.macro rgbconv16 dst, b, g, r, bc, gc, rc, shr_bits
         smull           v3.4s, \b\().4h, \bc
         smlal           v3.4s, \g\().4h, \gc
         smlal           v3.4s, \r\().4h, \rc
         smull2          v4.4s, \b\().8h, \bc
         smlal2          v4.4s, \g\().8h, \gc
         smlal2          v4.4s, \r\().8h, \rc        // v3:v4 = b * bc + g * gc + r * rc (32-bit)
-        shrn            \dst\().4h, v3.4s, #7
-        shrn2           \dst\().8h, v4.4s, #7       // dst = b * bc + g * gc + r * rc (16-bit)
+        shrn            \dst\().4h, v3.4s, \shr_bits
+        shrn2           \dst\().8h, v4.4s, \shr_bits       // dst = b * bc + g * gc + r * rc (16-bit)
 .endm
 
 // void ff_rgb24toyv12_neon(const uint8_t *src, uint8_t *ydst, uint8_t *udst,
@@ -171,8 +172,8 @@ function ff_rgb24toyv12_neon, export=1
         uxtl2           v21.8h, v28.16b             // v21 = R12
 
         // calculate Y values for first line
-        rgbconv16       v24, v16, v17, v18, BY, GY, RY // v24 = Y11
-        rgbconv16       v25, v19, v20, v21, BY, GY, RY // v25 = Y12
+        rgbconv16       v24, v16, v17, v18, BY, GY, RY, #7 // v24 = Y11
+        rgbconv16       v25, v19, v20, v21, BY, GY, RY, #7 // v25 = Y12
 
         // load second line
         ld3             {v26.16b, v27.16b, v28.16b}, [x10], #48
@@ -191,25 +192,22 @@ function ff_rgb24toyv12_neon, export=1
         uxtl2           v21.8h, v28.16b             // v21 = R22
 
         // calculate Y values for second line
-        rgbconv16       v26, v16, v17, v18, BY, GY, RY // v26 = Y21
-        rgbconv16       v27, v19, v20, v21, BY, GY, RY // v27 = Y22
+        rgbconv16       v26, v16, v17, v18, BY, GY, RY, #7 // v26 = Y21
+        rgbconv16       v27, v19, v20, v21, BY, GY, RY, #7 // v27 = Y22
 
         // pairwise add rgb values to calculate average
         addp            v16.8h, v16.8h, v19.8h
         addp            v17.8h, v17.8h, v20.8h
         addp            v18.8h, v18.8h, v21.8h
 
-        // calculate average
+        // calculate sum of r, g, b components in 2x2 blocks
         add             v16.8h, v16.8h, v5.8h
         add             v17.8h, v17.8h, v6.8h
         add             v18.8h, v18.8h, v7.8h
-        ushr            v16.8h, v16.8h, #2
-        ushr            v17.8h, v17.8h, #2
-        ushr            v18.8h, v18.8h, #2
 
         // calculate U and V values
-        rgbconv16       v28, v16, v17, v18, BU, GU, RU // v28 = U
-        rgbconv16       v29, v16, v17, v18, BV, GV, RV // v29 = V
+        rgbconv16       v28, v16, v17, v18, BU, GU, RU, #9 // v28 = U
+        rgbconv16       v29, v16, v17, v18, BV, GV, RV, #9 // v29 = V
 
         // add offsets and narrow all values
         addhn           v24.8b, v24.8h, Y_OFFSET.8h
-- 
2.47.2

_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel

To unsubscribe, visit link above, or email
ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".