[FFmpeg-cvslog] lavc/h264dsp: use saturing add/sub for R-V V 8-bit DC add

[Rémi Denis-Courmont]

ffmpeg | branch: master | Rémi Denis-Courmont <remi at remlab.net> | Thu Jul 25 20:47:34 2024 +0300| [b0b3bea10bab54f22f976245da343fb42c9c1f28] | committer: Rémi Denis-Courmont

lavc/h264dsp: use saturing add/sub for R-V V 8-bit DC add

T-Head C908 (cycles):
h264_idct4_dc_add_8bpp_c:      109.2
h264_idct4_dc_add_8bpp_rvv_i32: 34.5 (before)
h264_idct4_dc_add_8bpp_rvv_i32: 25.5 (after)
h264_idct8_dc_add_8bpp_c:      418.7
h264_idct8_dc_add_8bpp_rvv_i64: 69.5 (before)
h264_idct8_dc_add_8bpp_rvv_i64: 33.5 (after)

> http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=b0b3bea10bab54f22f976245da343fb42c9c1f28
---

 libavcodec/riscv/h264dsp_init.c |  5 +++--
 libavcodec/riscv/h264idct_rvv.S | 28 ++++++++++++++++++----------
 2 files changed, 21 insertions(+), 12 deletions(-)

diff --git a/libavcodec/riscv/h264dsp_init.c b/libavcodec/riscv/h264dsp_init.c
index 836c073559..671330d664 100644
--- a/libavcodec/riscv/h264dsp_init.c
+++ b/libavcodec/riscv/h264dsp_init.c
@@ -98,8 +98,8 @@ av_cold void ff_h264dsp_init_riscv(H264DSPContext *dsp, const int bit_depth,
 
             dsp->h264_idct_add  = ff_h264_idct_add_8_rvv;
             dsp->h264_idct8_add = ff_h264_idct8_add_8_rvv;
-            dsp->h264_idct_dc_add = ff_h264_idct4_dc_add_8_rvv;
             if (flags & AV_CPU_FLAG_RVB) {
+                dsp->h264_idct_dc_add     = ff_h264_idct4_dc_add_8_rvv;
                 dsp->h264_idct_add16      = ff_h264_idct_add16_8_rvv;
                 dsp->h264_idct_add16intra = ff_h264_idct_add16intra_8_rvv;
 #  if __riscv_xlen == 64
@@ -108,7 +108,8 @@ av_cold void ff_h264dsp_init_riscv(H264DSPContext *dsp, const int bit_depth,
             }
             if (flags & AV_CPU_FLAG_RVV_I64) {
                 dsp->h264_add_pixels8_clear = ff_h264_add_pixels8_8_rvv;
-                dsp->h264_idct8_dc_add      = ff_h264_idct8_dc_add_8_rvv;
+                if (flags & AV_CPU_FLAG_RVB)
+                    dsp->h264_idct8_dc_add = ff_h264_idct8_dc_add_8_rvv;
             }
             dsp->h264_add_pixels4_clear = ff_h264_add_pixels4_8_rvv;
         }
diff --git a/libavcodec/riscv/h264idct_rvv.S b/libavcodec/riscv/h264idct_rvv.S
index 5dd92de77d..f823346c8d 100644
--- a/libavcodec/riscv/h264idct_rvv.S
+++ b/libavcodec/riscv/h264idct_rvv.S
@@ -420,7 +420,7 @@ endfunc
 .endr
 
 .macro idct_dc_add8 width
-func ff_h264_idct\width\()_dc_add_8_rvv, zve64x
+func ff_h264_idct\width\()_dc_add_8_rvv, zve64x, b
         lpad    0
 .if \width == 8
         vsetivli        zero, \width, e8, mf2, ta, ma
@@ -428,26 +428,34 @@ func ff_h264_idct\width\()_dc_add_8_rvv, zve64x
         vsetivli        zero, \width, e8, mf4, ta, ma
 .endif
         lh              t0, 0(a1)
+        li              t1, 255
         addi            t0, t0, 32
         srai            t0, t0, 6
         sh              zero, 0(a1)
 .if \width == 8
         li              a6, \width * \width
         vlse64.v        v24, (a0), a2
-        vsetvli         zero, a6, e16, m8, ta, ma
+        vsetvli         zero, a6, e8, m4, ta, ma
 .else
         vlse32.v        v24, (a0), a2
-        vsetivli        zero, \width * \width, e16, m2, ta, ma
+        vsetivli        zero, \width * \width, e8, m1, ta, ma
 .endif
-        vzext.vf2       v0, v24
-        vadd.vx         v0, v0, t0
-        vmax.vx         v0, v0, zero
-.if \width == 8
-        vsetvli         zero, zero, e8, m4, ta, ma
+        bgez            t0, 1f
+
+        neg             t0, t0
+        minu            t0, t0, t1
+        vssubu.vx       v24, v24, t0
+ .if \width == 8
+        vsetivli        zero, \width, e8, mf2, ta, ma
+        vsse64.v        v24, (a0), a2
 .else
-        vsetvli         zero, zero, e8, m1, ta, ma
+        vsetivli        zero, \width, e8, mf4, ta, ma
+        vsse32.v        v24, (a0), a2
 .endif
-        vnclipu.wi      v24, v0, 0
+        ret
+1:
+        minu            t0, t0, t1
+        vsaddu.vx       v24, v24, t0
 .if \width == 8
         vsetivli        zero, \width, e8, mf2, ta, ma
         vsse64.v        v24, (a0), a2