diff mbox series

[FFmpeg-devel,4/4] lavu/fixed_dsp: R-V V vector_fmul_window

Message ID 20231004195110.38615-4-remi@remlab.net
State Accepted
Commit f39a8790e165589d0498ea86958d77b2b9fab8e8
Headers show
Series [FFmpeg-devel,1/4] lavu/fixed_dsp: R-V V vector_fmul_add | expand

Checks

Context Check Description
yinshiyou/configure_loongarch64 warning Failed to apply patch

Commit Message

Rémi Denis-Courmont Oct. 4, 2023, 7:51 p.m. UTC
---
 libavutil/riscv/fixed_dsp_init.c |  6 +++++
 libavutil/riscv/fixed_dsp_rvv.S  | 44 ++++++++++++++++++++++++++++++++
 2 files changed, 50 insertions(+)

Comments

Rémi Denis-Courmont Oct. 5, 2023, 2:54 p.m. UTC | #1
Le keskiviikkona 4. lokakuuta 2023, 22.51.10 EEST Rémi Denis-Courmont a écrit 
:
> +        vwmul.vv v12, v16, v28
> +        vwmul.vv v8, v16, v24
> +        // vwnmsac.vv does _not_ exist so multiply & subtract separately
> +        vwmul.vv v4, v20, v24
> +        vwmacc.vv v8, v20, v28

Retaining 3 whole double-width multiply vectors is really hurting here. This 
would be a lot smoother with just the top-halves (vsmul.vv), but then we need 
to round each multiplication. Of course the final results would differ slightly 
then, and I am not sure what that entails for the audio codecs (nor would I 
know how to update the x86 code).

N.B.: Don't ask for benchmarks, this can't run on v0.7.1 proto-RVV hardware.
diff mbox series

Patch

diff --git a/libavutil/riscv/fixed_dsp_init.c b/libavutil/riscv/fixed_dsp_init.c
index 02883b31d4..6469b45374 100644
--- a/libavutil/riscv/fixed_dsp_init.c
+++ b/libavutil/riscv/fixed_dsp_init.c
@@ -25,6 +25,9 @@ 
 #include "libavutil/cpu.h"
 #include "libavutil/fixed_dsp.h"
 
+void ff_vector_fmul_window_fixed_rvv(int32_t *dst, const int32_t *src0,
+                                     const int32_t *src1, const int32_t *win,
+                                     int len);
 void ff_vector_fmul_fixed_rvv(int *dst, const int *src0, const int *src1,
                               int len);
 void ff_vector_fmul_reverse_fixed_rvv(int *dst, const int *src0,
@@ -40,6 +43,9 @@  av_cold void ff_fixed_dsp_init_riscv(AVFixedDSPContext *fdsp)
     int flags = av_get_cpu_flags();
 
     if ((flags & AV_CPU_FLAG_RVV_I32) && (flags & AV_CPU_FLAG_RVB_ADDR)) {
+        if (flags & AV_CPU_FLAG_RVV_I64)
+            fdsp->vector_fmul_window = ff_vector_fmul_window_fixed_rvv;
+
         fdsp->vector_fmul = ff_vector_fmul_fixed_rvv;
         fdsp->vector_fmul_reverse = ff_vector_fmul_reverse_fixed_rvv;
         fdsp->vector_fmul_add = ff_vector_fmul_add_fixed_rvv;
diff --git a/libavutil/riscv/fixed_dsp_rvv.S b/libavutil/riscv/fixed_dsp_rvv.S
index d5d538239f..5b666016a0 100644
--- a/libavutil/riscv/fixed_dsp_rvv.S
+++ b/libavutil/riscv/fixed_dsp_rvv.S
@@ -20,6 +20,50 @@ 
 
 #include "asm.S"
 
+func ff_vector_fmul_window_fixed_rvv, zve64x
+        csrwi   vxrm, 0
+        vsetvli t0, zero, e16, m1, ta, ma
+        sh2add  a2, a4, a2
+        vid.v   v0
+        sh3add  t3, a4, a3
+        vadd.vi v0, v0, 1
+        sh3add  t0, a4, a0
+1:
+        vsetvli t2, a4, e16, m1, ta, ma
+        slli    t4, t2, 2
+        vrsub.vx v2, v0, t2
+        sub     t3, t3, t4
+        vsetvli zero, zero, e32, m2, ta, ma
+        sub     a2, a2, t4
+        vle32.v v8, (t3)
+        sub     t0, t0, t4
+        vle32.v v4, (a2)
+        sub     a4, a4, t2
+        vrgatherei16.vv v28, v8, v2
+        vle32.v v16, (a1)
+        add     a1, a1, t4
+        vrgatherei16.vv v20, v4, v2
+        vle32.v v24, (a3)
+        add     a3, a3, t4
+        vwmul.vv v12, v16, v28
+        vwmul.vv v8, v16, v24
+        // vwnmsac.vv does _not_ exist so multiply & subtract separately
+        vwmul.vv v4, v20, v24
+        vwmacc.vv v8, v20, v28
+        vsetvli zero, zero, e64, m4, ta, ma
+        vsub.vv v12, v12, v4
+        vsetvli zero, zero, e32, m2, ta, ma
+        vnclip.wi v16, v8, 31
+        vnclip.wi v20, v12, 31
+        vrgatherei16.vv v8, v16, v2
+        vse32.v v20, (a0)
+        add     a0, a0, t4
+        vse32.v v8, (t0)
+        bnez    a4, 1b
+
+        ret
+endfunc
+
 func ff_vector_fmul_fixed_rvv, zve32x
         csrwi   vxrm, 0
 1: