[FFmpeg-devel] [PATCH 4/4] lavc/h263dsp: R-V V dct_unquantize

[Rémi Denis-Courmont]

T-Head C908:
h263dsp.dct_unquantize_inter_c:       5.7
h263dsp.dct_unquantize_inter_rvv_i32: 2.7
h263dsp.dct_unquantize_intra_c:       5.7
h263dsp.dct_unquantize_intra_rvv_i32: 3.0

SpacemiT X60:
h263dsp.dct_unquantize_inter_c:       5.0
h263dsp.dct_unquantize_inter_rvv_i32: 1.2
h263dsp.dct_unquantize_intra_c:       5.0
h263dsp.dct_unquantize_intra_rvv_i32: 1.5
---
 libavcodec/riscv/h263dsp_init.c | 12 +++++++++---
 libavcodec/riscv/h263dsp_rvv.S  | 22 ++++++++++++++++++++++
 2 files changed, 31 insertions(+), 3 deletions(-)

diff --git a/libavcodec/riscv/h263dsp_init.c b/libavcodec/riscv/h263dsp_init.c
index 21b536366c..5d73fde865 100644
--- a/libavcodec/riscv/h263dsp_init.c
+++ b/libavcodec/riscv/h263dsp_init.c
@@ -25,6 +25,7 @@
 #include "libavutil/riscv/cpu.h"
 #include "libavcodec/h263dsp.h"
 
+void ff_h263_dct_unquantize_rvv(int16_t *, size_t start, size_t end, int, int);
 void ff_h263_h_loop_filter_rvv(uint8_t *src, int stride, int q);
 void ff_h263_v_loop_filter_rvv(uint8_t *src, int stride, int q);
 
@@ -33,9 +34,14 @@ av_cold void ff_h263dsp_init_riscv(H263DSPContext *c)
 #if HAVE_RVV
     int flags = av_get_cpu_flags();
 
-    if ((flags & AV_CPU_FLAG_RVV_I32) && ff_rv_vlen_least(128)) {
-        c->h263_h_loop_filter = ff_h263_h_loop_filter_rvv;
-        c->h263_v_loop_filter = ff_h263_v_loop_filter_rvv;
+    if (flags & AV_CPU_FLAG_RVV_I32) {
+        if (flags & AV_CPU_FLAG_RVB_ADDR)
+            c->h263_dct_unquantize = ff_h263_dct_unquantize_rvv;
+
+        if (ff_rv_vlen_least(128)) {
+            c->h263_h_loop_filter = ff_h263_h_loop_filter_rvv;
+            c->h263_v_loop_filter = ff_h263_v_loop_filter_rvv;
+        }
     }
 #endif
 }
diff --git a/libavcodec/riscv/h263dsp_rvv.S b/libavcodec/riscv/h263dsp_rvv.S
index 97503d527c..319c51a0bb 100644
--- a/libavcodec/riscv/h263dsp_rvv.S
+++ b/libavcodec/riscv/h263dsp_rvv.S
@@ -20,6 +20,28 @@
 
 #include "libavutil/riscv/asm.S"
 
+func ff_h263_dct_unquantize_rvv, zve32x
+        sub     a2, a2, a1
+        sh1add  a0, a1, a0
+        addi    a2, a2, 1
+1:
+        vsetvli  t0, a2, e16, m4, ta, mu
+        vle16.v  v8, (a0)
+        sub      a2, a2, t0
+        vmv.v.x  v24, a4
+        vmslt.vi v0, v8, 0
+        vneg.v   v24, v24, v0.t
+        vmsne.vi v0, v8, 0
+        vwmul.vx v16, v8, a3
+        vwadd.wv v16, v16, v24, v0.t
+        vncvt.x.x.w v8, v16
+        vse16.v v8, (a0)
+        sh1add  a0, t0, a0
+        bnez    a2, 1b
+
+        ret
+endfunc
+
         .option push
         .option norelax
 func ff_h263_h_loop_filter_rvv, zve32x
-- 
2.45.1