[FFmpeg-devel] [PATCH 2/2] lavc/aacpsdsp: rework R-V V hybrid_synthesis_deint

[Rémi Denis-Courmont]

Given the size of the data set, strided memory accesses cannot be avoided.
We can still do better than the current code.

ps_hybrid_synthesis_deint_c:       12065.5
ps_hybrid_synthesis_deint_rvv_i32: 13650.2 (before)
ps_hybrid_synthesis_deint_rvv_i64:  8181.0 (after)
---
 libavcodec/riscv/aacpsdsp_init.c |  8 ++---
 libavcodec/riscv/aacpsdsp_rvv.S  | 61 +++++++++++++++++---------------
 2 files changed, 36 insertions(+), 33 deletions(-)

diff --git a/libavcodec/riscv/aacpsdsp_init.c b/libavcodec/riscv/aacpsdsp_init.c
index f72d1bc330..e094660cf3 100644
--- a/libavcodec/riscv/aacpsdsp_init.c
+++ b/libavcodec/riscv/aacpsdsp_init.c
@@ -46,16 +46,16 @@ av_cold void ff_psdsp_init_riscv(PSDSPContext *c)
         c->hybrid_analysis = ff_ps_hybrid_analysis_rvv;
 
         if (flags & AV_CPU_FLAG_RVB_ADDR) {
-            if (flags & AV_CPU_FLAG_RVV_I64)
+            if (flags & AV_CPU_FLAG_RVV_I64) {
                 c->add_squares = ff_ps_add_squares_rvv;
+                c->hybrid_synthesis_deint = ff_ps_hybrid_synthesis_deint_rvv;
+            }
             c->mul_pair_single = ff_ps_mul_pair_single_rvv;
             c->stereo_interpolate[0] = ff_ps_stereo_interpolate_rvv;
         }
     }
 
-    if ((flags & AV_CPU_FLAG_RVV_I32) && (flags & AV_CPU_FLAG_RVB_ADDR)) {
+    if ((flags & AV_CPU_FLAG_RVV_I32) && (flags & AV_CPU_FLAG_RVB_ADDR))
         c->hybrid_analysis_ileave = ff_ps_hybrid_analysis_ileave_rvv;
-        c->hybrid_synthesis_deint = ff_ps_hybrid_synthesis_deint_rvv;
-    }
 #endif
 }
diff --git a/libavcodec/riscv/aacpsdsp_rvv.S b/libavcodec/riscv/aacpsdsp_rvv.S
index cf872599c8..1dc426e01c 100644
--- a/libavcodec/riscv/aacpsdsp_rvv.S
+++ b/libavcodec/riscv/aacpsdsp_rvv.S
@@ -190,38 +190,41 @@ func ff_ps_hybrid_analysis_ileave_rvv, zve32x /* no needs for zve32f here */
         ret
 endfunc
 
-func ff_ps_hybrid_synthesis_deint_rvv, zve32x
-        slli        t1, a2, 5 + 1 + 2
-        sh2add      a0, a2, a0
-        add         a1, a1, t1
-        addi        a2, a2, -64
-        li          t1, 38 * 64 * 4
-        li          t6, 64 * 4
-        add         a4, a0, t1
-        beqz        a2, 3f
+func ff_ps_hybrid_synthesis_deint_rvv, zve64x
+        slli    t0, a2, 5 + 1 + 2
+        sh2add  a0, a2, a0
+        add     a1, a1, t0
+        addi    t2, a2, -64
+        li      t0, 38 * 64
+        li      t1, 32 * 2 * 4
+        li      t4, 8 - 16384 // offset from in[64][n][0] to in[0][n + 1][0]
+        slli    t5, a2, 5 + 1 + 2 // and from in[0][n+1][0] to in[0][n+1][s]
+        neg     t2, t2
+        li      t3, 32
+        add     a4, t4, t5
+        sh2add  t0, t0, a0
 1:
-        mv          t0, a0
-        mv          t1, a1
-        mv          t3, a3
-        mv          t4, a4
-        addi        a2, a2, 1
+        mv      t4, t2
+        addi    a3, a3, -1
 2:
-        vsetvli     t5, t3, e32, m4, ta, ma
-        vlseg2e32.v v16, (t1)
-        sub         t3, t3, t5
-        vsse32.v    v16, (t0), t6
-        mul         t2, t5, t6
-        vsse32.v    v20, (t4), t6
-        sh3add      t1, t5, t1
-        add         t0, t0, t2
-        add         t4, t4, t2
-        bnez        t3, 2b
+        vsetvli t5, t4, e32, m4, ta, ma
+        vlse64.v v16, (a1), t1 /* sizeof (float[32][2]) */
+        sub     t4, t4, t5
+        vnsrl.wx v24, v16, zero
+        slli    t6, t5, 5 + 1 + 2
+        vnsrl.wx v28, v16, t3 /* 32 */
+        add     a1, a1, t6
+        vse32.v v24, (a0)
+        sh2add  a0, t5, a0
+        vse32.v v28, (t0)
+        sh2add  t0, t5, t0
+        bnez    t4, 2b
+
+        add     a1, a1, a4
+        sh2add  a0, a2, a0
+        sh2add  t0, a2, t0
+        bnez    a3, 1b
 
-        add         a0, a0, 4
-        add         a1, a1, 32 * 2 * 4
-        add         a4, a4, 4
-        bnez        a2, 1b
-3:
         ret
 endfunc
 
-- 
2.42.0