[FFmpeg-devel] [PATCH 2/3] swscale/rgb2rgb2: rework RISC-V V shuffle_bytes_{0321, 2103}

[Rémi Denis-Courmont]

This avoids strided loads.

Before:
shuffle_bytes_0321_rvv_i32: 307.7
shuffle_bytes_2103_rvv_i32: 308.7

After:
shuffle_bytes_0321_rvv_i32: 59.7
shuffle_bytes_2103_rvv_i32: 61.5
---
 libswscale/riscv/rgb2rgb_rvv.S | 45 +++++++++++++++-------------------
 1 file changed, 20 insertions(+), 25 deletions(-)

diff --git a/libswscale/riscv/rgb2rgb_rvv.S b/libswscale/riscv/rgb2rgb_rvv.S
index 8f30d760cb..366f80c1c3 100644
--- a/libswscale/riscv/rgb2rgb_rvv.S
+++ b/libswscale/riscv/rgb2rgb_rvv.S
@@ -21,38 +21,33 @@
 #include "libavutil/riscv/asm.S"
 
 func ff_shuffle_bytes_0321_rvv, zve32x
-        addi    t1, a0, 3
-        addi    t2, a0, 2
-        addi    t3, a0, 1
+        li      t1, 0x00ff00ff
+        j       1f
+endfunc
+
+func ff_shuffle_bytes_2103_rvv, zve32x
+        li      t1, ~0x00ff00ff
 1:
+        not     t2, t1
         srai    a2, a2, 2
-        li      t4, 4
 2:
-        vsetvli    t0, a2, e8, m1, ta, ma
-        sub        a2, a2, t0
-        vlse8.v    v8, (a0), t4
-        sh2add     a0, t0, a0
-        vlse8.v    v9, (t1), t4
-        sh2add     t1, t0, t1
-        vlse8.v    v10, (t2), t4
-        sh2add     t2, t0, t2
-        vlse8.v    v11, (t3), t4
-        sh2add     t3, t0, t3
-        vsseg4e8.v v8, (a1)
-        sh2add     a1, t0, a1
-        bnez       a2, 2b
+        vsetvli t0, a2, e32, m8, ta, ma
+        vle32.v v8, (a0)
+        sub     a2, a2, t0
+        vand.vx v16, v8, t2
+        sh2add  a0, t0, a0
+        vand.vx v8, v8, t1
+        vsrl.vi v24, v16, 16
+        vsll.vi v16, v16, 16
+        vor.vv  v8, v8, v24
+        vor.vv  v8, v16, v8
+        vse32.v v8, (a1)
+        sh2add  a1, t0, a1
+        bnez    a2, 2b
 
         ret
 endfunc
 
-func ff_shuffle_bytes_2103_rvv, zve32x
-        addi    t1, a0, 1
-        addi    t2, a0, 0
-        addi    t3, a0, 3
-        addi    a0, a0, 2
-        j       1b
-endfunc
-
 func ff_shuffle_bytes_1230_rvv, zve32x
         addi    t1, a0, 2
         addi    t2, a0, 3
-- 
2.40.1