[FFmpeg-devel] [PATCH 2/2] lavc/mpegvideoencdsp: R-V V add_8x8basis

[Rémi Denis-Courmont]

T-Head C908:
add_8x8basis_c:      440.6
add_8x8basis_rvv_i32: 70.3

SpacemiT X60:
add_8x8basis_c:      436.3
add_8x8basis_rvv_i32: 40.5
---
 libavcodec/riscv/mpegvideoencdsp_init.c |  5 ++++-
 libavcodec/riscv/mpegvideoencdsp_rvv.S  | 19 +++++++++++++++++++
 2 files changed, 23 insertions(+), 1 deletion(-)

diff --git a/libavcodec/riscv/mpegvideoencdsp_init.c b/libavcodec/riscv/mpegvideoencdsp_init.c
index 4c156c1cf2..1ac808af16 100644
--- a/libavcodec/riscv/mpegvideoencdsp_init.c
+++ b/libavcodec/riscv/mpegvideoencdsp_init.c
@@ -25,6 +25,7 @@
 
 int ff_try_8x8basis_rvv(const int16_t rem[64], const int16_t weight[64],
                         const int16_t basis[16], int scale);
+void ff_add_8x8basis_rvv(int16_t rem[64], const int16_t basis[16], int scale);
 int ff_pix_sum_rvv(const uint8_t *pix, int line_size);
 int ff_pix_norm1_rvv(const uint8_t *pix, int line_size);
 
@@ -35,8 +36,10 @@ av_cold void ff_mpegvideoencdsp_init_riscv(MpegvideoEncDSPContext *c,
     int flags = av_get_cpu_flags();
 
     if (flags & AV_CPU_FLAG_RVV_I32) {
-        if (flags & AV_CPU_FLAG_RVB)
+        if (flags & AV_CPU_FLAG_RVB) {
             c->try_8x8basis = ff_try_8x8basis_rvv;
+            c->add_8x8basis = ff_add_8x8basis_rvv;
+        }
 
         if (flags & AV_CPU_FLAG_RVV_I64) {
             if ((flags & AV_CPU_FLAG_RVB) && ff_rv_vlen_least(128))
diff --git a/libavcodec/riscv/mpegvideoencdsp_rvv.S b/libavcodec/riscv/mpegvideoencdsp_rvv.S
index 9408de47c8..7c50526934 100644
--- a/libavcodec/riscv/mpegvideoencdsp_rvv.S
+++ b/libavcodec/riscv/mpegvideoencdsp_rvv.S
@@ -55,6 +55,25 @@ func ff_try_8x8basis_rvv, zve32x, b
         ret
 endfunc
 
+func ff_add_8x8basis_rvv, zve32x, b
+        li      t1, 64
+        csrwi   vxrm, 0
+1:
+        vsetvli     t0, t1, e16, m4, ta, ma
+        vle16.v     v4, (a1)
+        sub     t1, t1, t0
+        vwmul.vx    v16, v4, a2
+        sh1add  a1, t0, a1
+        vle16.v     v8, (a0)
+        vnclip.wi   v4, v16, BASIS_SHIFT - RECON_SHIFT
+        vadd.vv     v4, v8, v4
+        vse16.v     v4, (a0)
+        sh1add  a0, t0, a0
+        bnez    t1, 1b
+
+        ret
+endfunc
+
 func ff_pix_sum_rvv, zve64x, b
         lpad    0
         vsetivli    t0, 16, e16, m1, ta, ma
-- 
2.45.2