[FFmpeg-devel] [PATCH] lavc/h264dsp: stick R-V V biweight to 16-bit

[Rémi Denis-Courmont]

T-Head C908 (ns):
h264_biweight2_8_c:        2414.5
h264_biweight2_8_rvv_i32:   701.8 (before)
h264_biweight2_8_rvv_i32:   562.8 (after)
h264_biweight4_8_c:        4655.3
h264_biweight4_8_rvv_i32:  1377.5 (before)
h264_biweight4_8_rvv_i32:  1109.0 (after)
h264_biweight8_8_c:        9701.5
h264_biweight8_8_rvv_i32:  2896.0 (before)
h264_biweight8_8_rvv_i32:  2340.5 (after)
h264_biweight16_8_c:      18025.0
h264_biweight16_8_rvv_i32: 3460.8 (before)
h264_biweight16_8_rvv_i32: 3146.0 (after)

SpacemiT X60 (ns):
h264_biweight2_8_c:        2415.5
h264_biweight2_8_rvv_i32:   478.2 (before)
h264_biweight2_8_rvv_i32:   405.0 (after)
h264_biweight4_8_c:        4655.3
h264_biweight4_8_rvv_i32:   946.7 (before)
h264_biweight4_8_rvv_i32:   811.5 (after)
h264_biweight8_8_c:        9061.8
h264_biweight8_8_rvv_i32:  2071.7 (before)
h264_biweight8_8_rvv_i32:  1811.3 (after)
h264_biweight16_8_c:      18020.5
h264_biweight16_8_rvv_i32: 3457.2 (before)
h264_biweight16_8_rvv_i32: 2499.0 (after)
---
 libavcodec/riscv/h264dsp_rvv.S | 32 +++++++++++++++-----------------
 1 file changed, 15 insertions(+), 17 deletions(-)

diff --git a/libavcodec/riscv/h264dsp_rvv.S b/libavcodec/riscv/h264dsp_rvv.S
index a38bf7ef1d..46392c92a5 100644
--- a/libavcodec/riscv/h264dsp_rvv.S
+++ b/libavcodec/riscv/h264dsp_rvv.S
@@ -56,22 +56,21 @@ func ff_h264_biweight_pixels_simple_8_rvv, zve32x
         addi    a7, a7, 1
         ori     a7, a7, 1
         sll     a7, a7, a4
+        addi    a4, a4, 1
 1:
-        vsetvli zero, t6, e32, m4, ta, ma
+        vsetvli zero, t6, e16, m2, ta, ma
         vle8.v  v8, (a0)
         addi    a3, a3, -1
         vle8.v  v12, (a1)
         add     a1, a1, a2
         vmv.v.x v16, a7
-        vsetvli zero, zero, e16, m2, ta, ma
         vzext.vf2   v24, v8
         vzext.vf2   v28, v12
-        vwmaccsu.vx v16, a5, v24
-        vwmaccsu.vx v16, a6, v28
-        vnclip.wx   v16, v16, a4
+        vmacc.vx    v16, a5, v24
+        vmacc.vx    v16, a6, v28
         vmax.vx v16, v16, zero
         vsetvli zero, zero, e8, m1, ta, ma
-        vnclipu.wi  v8, v16, 1
+        vnclipu.wx  v8, v16, a4
         vse8.v  v8, (a0)
         add     a0, a0, a2
         bnez    a3, 1b
@@ -121,12 +120,13 @@ func ff_h264_biweight_pixels_8_rvv, zve32x
         addi    a7, a7, 1
         ori     a7, a7, 1
         sll     a7, a7, a4
+        addi    a4, a4, 1
 1:
         mv      t0, a0
         mv      t1, a1
         mv      t5, t6
 2:
-        vsetvli t2, a3, e32, m8, ta, ma
+        vsetvli t2, a3, e16, m4, ta, ma
         vlsseg2e8.v v0, (t0), a2
         vlsseg2e8.v v4, (t1), a2
         addi    t5, t5, -2
@@ -135,19 +135,17 @@ func ff_h264_biweight_pixels_8_rvv, zve32x
         vsetvli zero, zero, e16, m4, ta, ma
         vzext.vf2   v8, v0
         vzext.vf2   v12, v2
-        vwmaccsu.vx v16, a5, v8
-        vwmaccsu.vx v24, a5, v12
+        vmacc.vx    v16, a5, v8
+        vmacc.vx    v24, a5, v12
         vzext.vf2   v8, v4
         vzext.vf2   v12, v6
-        vwmaccsu.vx v16, a6, v8
-        vwmaccsu.vx v24, a6, v12
-        vnclip.wx   v8, v16, a4
-        vnclip.wx   v12, v24, a4
-        vmax.vx v8, v8, zero
-        vmax.vx v12, v12, zero
+        vmacc.vx    v16, a6, v8
+        vmacc.vx    v24, a6, v12
+        vmax.vx     v8, v16, zero
+        vmax.vx     v12, v24, zero
         vsetvli zero, zero, e8, m2, ta, ma
-        vnclipu.wi  v0, v8, 1
-        vnclipu.wi  v2, v12, 1
+        vnclipu.wx  v0, v8, a4
+        vnclipu.wx  v2, v12, a4
         vssseg2e8.v v0, (t0), a2
         addi    t0, t0, 2
         addi    t1, t1, 2
-- 
2.45.2