[FFmpeg-devel] [PATCH 2/2] lavc/vc1dsp: R-V V vc1_inv_trans_8x4

[Rémi Denis-Courmont]

T-Head C908:
vc1dsp.vc1_inv_trans_8x4_c:      10.5
vc1dsp.vc1_inv_trans_8x4_rvv_i32: 3.5
---
 libavcodec/riscv/vc1dsp_init.c |  2 +
 libavcodec/riscv/vc1dsp_rvv.S  | 74 ++++++++++++++++++++++++++++++++++
 2 files changed, 76 insertions(+)

diff --git a/libavcodec/riscv/vc1dsp_init.c b/libavcodec/riscv/vc1dsp_init.c
index b8a1015ce5..e63870ad44 100644
--- a/libavcodec/riscv/vc1dsp_init.c
+++ b/libavcodec/riscv/vc1dsp_init.c
@@ -29,6 +29,7 @@ void ff_vc1_inv_trans_8x8_dc_rvv(uint8_t *dest, ptrdiff_t stride, int16_t *block
 void ff_vc1_inv_trans_8x8_rvv(int16_t block[64]);
 void ff_vc1_inv_trans_4x8_dc_rvv(uint8_t *dest, ptrdiff_t stride, int16_t *block);
 void ff_vc1_inv_trans_8x4_dc_rvv(uint8_t *dest, ptrdiff_t stride, int16_t *block);
+void ff_vc1_inv_trans_8x4_rvv(uint8_t *dest, ptrdiff_t stride, int16_t *block);
 void ff_vc1_inv_trans_4x4_dc_rvv(uint8_t *dest, ptrdiff_t stride, int16_t *block);
 void ff_put_pixels16x16_rvi(uint8_t *dst, const uint8_t *src, ptrdiff_t line_size, int rnd);
 void ff_put_pixels8x8_rvi(uint8_t *dst, const uint8_t *src, ptrdiff_t line_size, int rnd);
@@ -55,6 +56,7 @@ av_cold void ff_vc1dsp_init_riscv(VC1DSPContext *dsp)
     if (flags & AV_CPU_FLAG_RVV_I32) {
         if (ff_rv_vlen_least(128)) {
             dsp->vc1_inv_trans_8x8 = ff_vc1_inv_trans_8x8_rvv;
+            dsp->vc1_inv_trans_8x4 = ff_vc1_inv_trans_8x4_rvv;
             dsp->vc1_inv_trans_4x8_dc = ff_vc1_inv_trans_4x8_dc_rvv;
             dsp->vc1_inv_trans_4x4_dc = ff_vc1_inv_trans_4x4_dc_rvv;
             dsp->avg_vc1_mspel_pixels_tab[0][0] = ff_avg_pixels16x16_rvv;
diff --git a/libavcodec/riscv/vc1dsp_rvv.S b/libavcodec/riscv/vc1dsp_rvv.S
index 0af4d26a11..7ac1062dcd 100644
--- a/libavcodec/riscv/vc1dsp_rvv.S
+++ b/libavcodec/riscv/vc1dsp_rvv.S
@@ -180,6 +180,29 @@ func ff_vc1_inv_trans_8_rvv, zve32x
         jr       t0
 endfunc
 
+        .variant_cc ff_vc1_inv_trans_4_rvv
+func ff_vc1_inv_trans_4_rvv, zve32x
+        li       t3, 17
+        vmul.vx  v8, v0, t3
+        li       t4, 22
+        vmul.vx  v10, v2, t3
+        li       t2, 10
+        vadd.vx  v8, v8, t1 # +4 or +64
+        vmul.vx  v14, v1, t4
+        vadd.vv  v24, v8, v10  # t1
+        vsub.vv  v25, v8, v10  # t2
+        vmul.vx  v16, v3, t2
+        vmul.vx  v18, v3, t4
+        vmul.vx  v20, v1, t2
+        vadd.vv  v26, v14, v16 # t3
+        vsub.vv  v27, v18, v20 # t4
+        vadd.vv  v0, v24, v26
+        vsub.vv  v1, v25, v27
+        vadd.vv  v2, v25, v27
+        vsub.vv  v3, v24, v26
+        jr       t0
+endfunc
+
 func ff_vc1_inv_trans_8x8_rvv, zve32x
         vsetivli zero, 8, e16, m1, ta, ma
         addi     a1, a0, 1 * 8 * 2
@@ -231,6 +254,57 @@ func ff_vc1_inv_trans_8x8_rvv, zve32x
         ret
 endfunc
 
+func ff_vc1_inv_trans_8x4_rvv, zve32x
+        vsetivli  zero, 4, e16, mf2, ta, ma
+        vlseg8e16.v v0, (a2)
+        li        t1, 4
+        jal       t0, ff_vc1_inv_trans_8_rvv
+        vsseg8e16.v v0, (a2)
+        addi      a3, a2, 1 * 8 * 2
+        vsetivli  zero, 8, e16, m1, ta, ma
+        vle16.v   v0, (a2)
+        addi      a4, a2, 2 * 8 * 2
+        vle16.v   v1, (a3)
+        addi      a5, a2, 3 * 8 * 2
+        vle16.v   v2, (a4)
+        li        t1, 64 - (128 << 7) # bias for signed vnclip.wi below
+        vle16.v   v3, (a5)
+        .irp n,0,1,2,3
+        # shift 4 vectors of 8 elems after transpose instead of 8 of 4
+        vsra.vi   v\n, v\n, 3
+        .endr
+        jal       t0, ff_vc1_inv_trans_4_rvv
+        add       a3, a1, a0
+        vle8.v    v8, (a0)
+        add       a4, a1, a3
+        vle8.v    v9, (a3)
+        add       a5, a1, a4
+        vle8.v    v10, (a4)
+        li        t1, 128
+        vle8.v    v11, (a5)
+        .irp n,0,1,2,3
+        vsra.vi   v\n, v\n, 7
+        .endr
+        vsetvli   zero, zero, e8, mf2, ta, ma
+        vwaddu.wv v0, v0, v8
+        vwaddu.wv v1, v1, v9
+        vwaddu.wv v2, v2, v10
+        vwaddu.wv v3, v3, v11
+        vnclip.wi v8, v0, 0
+        vnclip.wi v9, v1, 0
+        vnclip.wi v10, v2, 0
+        vnclip.wi v11, v3, 0
+        vxor.vx   v8, v8, t1
+        vxor.vx   v9, v9, t1
+        vse8.v    v8, (a0)
+        vxor.vx   v10, v10, t1
+        vse8.v    v9, (a3)
+        vxor.vx   v11, v11, t1
+        vse8.v    v10, (a4)
+        vse8.v    v11, (a5)
+        ret
+endfunc
+
 .macro mspel_op op pos n1 n2
         add           t1, \pos, a2
         v\op\()e8.v   v\n1, (\pos)
-- 
2.45.1