[FFmpeg-devel] [PATCH] lavu/float_dsp: unroll RISC-V V loops

[Rémi Denis-Courmont]

butterflies_float_c: 1057.0
butterflies_float_rvv_f32: 351.0 (before)
butterflies_float_rvv_f32: 329.5 (after)

vector_dmac_scalar_c: 819.0
vector_dmac_scalar_rvv_f64: 670.5 (before)
vector_dmac_scalar_rvv_f64: 431.0 (after)

vector_dmul_c: 800.2
vector_dmul_rvv_f64: 541.5 (before)
vector_dmul_rvv_f64: 426.0 (after)

vector_dmul_scalar_c: 545.7
vector_dmul_scalar_rvv_f64: 670.7 (before)
vector_dmul_scalar_rvv_f64: 324.7 (after)

vector_fmac_scalar_c: 804.5
vector_fmac_scalar_rvv_f32: 412.7 (before)
vector_fmac_scalar_rvv_f32: 214.5 (after)

vector_fmul_c: 811.2
vector_fmul_rvv_f32: 285.7 (before)
vector_fmul_rvv_f32: 214.2 (after)

vector_fmul_add_c: 1313.0
vector_fmul_add_rvv_f32: 349.0 (before)
vector_fmul_add_rvv_f32: 290.2 (after)

vector_fmul_reverse_c: 815.7
vector_fmul_reverse_rvv_f32: 529.2 (before)
vector_fmul_reverse_rvv_f32: 515.7 (after)

vector_fmul_scalar_c: 546.0
vector_fmul_scalar_rvv_f32: 350.2 (before)
vector_fmul_scalar_rvv_f32: 169.5 (after)
---
 libavutil/riscv/float_dsp_rvv.S | 20 ++++++++++----------
 1 file changed, 10 insertions(+), 10 deletions(-)

diff --git a/libavutil/riscv/float_dsp_rvv.S b/libavutil/riscv/float_dsp_rvv.S
index c80fde2f7e..bcec7074b8 100644
--- a/libavutil/riscv/float_dsp_rvv.S
+++ b/libavutil/riscv/float_dsp_rvv.S
@@ -23,7 +23,7 @@
 // (a0) = (a1) * (a2) [0..a3-1]
 func ff_vector_fmul_rvv, zve32f
 1:
-        vsetvli  t0, a3, e32, m1, ta, ma
+        vsetvli  t0, a3, e32, m8, ta, ma
         vle32.v  v16, (a1)
         sub      a3, a3, t0
         vle32.v  v24, (a2)
@@ -42,7 +42,7 @@ func ff_vector_fmac_scalar_rvv, zve32f
 NOHWF   fmv.w.x   fa0, a2
 NOHWF   mv        a2, a3
 1:
-        vsetvli   t0, a2, e32, m1, ta, ma
+        vsetvli   t0, a2, e32, m8, ta, ma
         slli      t1, t0, 2
         vle32.v   v24, (a1)
         sub       a2, a2, t0
@@ -61,7 +61,7 @@ func ff_vector_fmul_scalar_rvv, zve32f
 NOHWF   fmv.w.x  fa0, a2
 NOHWF   mv       a2, a3
 1:
-        vsetvli  t0, a2, e32, m1, ta, ma
+        vsetvli  t0, a2, e32, m8, ta, ma
         vle32.v  v16, (a1)
         sub      a2, a2, t0
         vfmul.vf v16, v16, fa0
@@ -82,7 +82,7 @@ func ff_vector_fmul_window_rvv, zve32f
         sh2add     t3, t1, a3
         li         t1, -4 // byte stride
 1:
-        vsetvli    t2, a4, e32, m1, ta, ma
+        vsetvli    t2, a4, e32, m4, ta, ma
         vle32.v    v16, (a1)
         slli       t4, t2, 2
         vlse32.v   v20, (a2), t1
@@ -109,7 +109,7 @@ endfunc
 // (a0) = (a1) * (a2) + (a3) [0..a4-1]
 func ff_vector_fmul_add_rvv, zve32f
 1:
-        vsetvli   t0, a4, e32, m1, ta, ma
+        vsetvli   t0, a4, e32, m8, ta, ma
         vle32.v   v8, (a1)
         sub       a4, a4, t0
         vle32.v   v16, (a2)
@@ -131,7 +131,7 @@ func ff_vector_fmul_reverse_rvv, zve32f
         li       t2, -4 // byte stride
         addi     a2, a2, -4
 1:
-        vsetvli  t0, a3, e32, m1, ta, ma
+        vsetvli  t0, a3, e32, m8, ta, ma
         slli     t1, t0, 2
         vle32.v  v16, (a1)
         sub      a3, a3, t0
@@ -149,7 +149,7 @@ endfunc
 // (a0) = (a0) + (a1), (a1) = (a0) - (a1) [0..a2-1]
 func ff_butterflies_float_rvv, zve32f
 1:
-        vsetvli  t0, a2, e32, m1, ta, ma
+        vsetvli  t0, a2, e32, m8, ta, ma
         vle32.v  v16, (a0)
         sub      a2, a2, t0
         vle32.v  v24, (a1)
@@ -188,7 +188,7 @@ endfunc
 // (a0) = (a1) * (a2) [0..a3-1]
 func ff_vector_dmul_rvv, zve64d
 1:
-        vsetvli  t0, a3, e64, m1, ta, ma
+        vsetvli  t0, a3, e64, m8, ta, ma
         vle64.v  v16, (a1)
         sub      a3, a3, t0
         vle64.v  v24, (a2)
@@ -207,7 +207,7 @@ func ff_vector_dmac_scalar_rvv, zve64d
 NOHWD   fmv.d.x   fa0, a2
 NOHWD   mv        a2, a3
 1:
-        vsetvli   t0, a2, e64, m1, ta, ma
+        vsetvli   t0, a2, e64, m8, ta, ma
         vle64.v   v24, (a1)
         sub       a2, a2, t0
         vle64.v   v16, (a0)
@@ -225,7 +225,7 @@ func ff_vector_dmul_scalar_rvv, zve64d
 NOHWD   fmv.d.x  fa0, a2
 NOHWD   mv       a2, a3
 1:
-        vsetvli  t0, a2, e64, m1, ta, ma
+        vsetvli  t0, a2, e64, m8, ta, ma
         vle64.v  v16, (a1)
         sub      a2, a2, t0
         vfmul.vf v16, v16, fa0
-- 
2.40.1