[FFmpeg-devel] [PATCHv2 4/4] lavc/vc1dsp: R-V V vc1_inv_trans_4x4

[Rémi Denis-Courmont]

T-Head C908 (cycles):
vc1dsp.vc1_inv_trans_4x4_c: 310.7
vc1dsp.vc1_inv_trans_4x4_rvv_i32: 120.0

We could use 1 `vlseg4e64.v` instead of 4 `vle16.v`, but that seems to
be about 7% slower.
---
 libavcodec/riscv/vc1dsp_init.c |  2 ++
 libavcodec/riscv/vc1dsp_rvv.S  | 47 ++++++++++++++++++++++++++++++++++
 2 files changed, 49 insertions(+)

diff --git a/libavcodec/riscv/vc1dsp_init.c b/libavcodec/riscv/vc1dsp_init.c
index cf9d42f377..de9002f395 100644
--- a/libavcodec/riscv/vc1dsp_init.c
+++ b/libavcodec/riscv/vc1dsp_init.c
@@ -32,6 +32,7 @@ void ff_vc1_inv_trans_4x8_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_vc1_inv_trans_4x4_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);
 void ff_avg_pixels16x16_rvv(uint8_t *dst, const uint8_t *src, ptrdiff_t line_size, int rnd);
@@ -59,6 +60,7 @@ av_cold void ff_vc1dsp_init_riscv(VC1DSPContext *dsp)
             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 = ff_vc1_inv_trans_4x8_rvv;
+            dsp->vc1_inv_trans_4x4 = ff_vc1_inv_trans_4x4_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 e22950d25d..03c942cba2 100644
--- a/libavcodec/riscv/vc1dsp_rvv.S
+++ b/libavcodec/riscv/vc1dsp_rvv.S
@@ -113,6 +113,8 @@ func ff_vc1_inv_trans_4x4_dc_rvv, zve32x
         ret
 endfunc
 
+#include "../../riscv-tests/debug.S"
+
         .variant_cc ff_vc1_inv_trans_8_rvv
 func ff_vc1_inv_trans_8_rvv, zve32x
         li      t4, 12
@@ -373,6 +375,51 @@ func ff_vc1_inv_trans_4x8_rvv, zve32x
         ret
 endfunc
 
+func ff_vc1_inv_trans_4x4_rvv, zve32x
+        li           a3, 8 * 2
+        csrwi        vxrm, 0
+        vsetivli     zero, 4, e16, mf2, ta, ma
+        vlsseg4e16.v v0, (a2), a3
+        li           t1, 3
+        jal          t0, ff_vc1_inv_trans_4_rvv
+        vsseg4e16.v  v0, (a2)
+        addi         t1, a2, 1 * 4 * 2
+        vle16.v      v0, (a2)
+        addi         t2, a2, 2 * 4 * 2
+        vle16.v      v1, (t1)
+        addi         t3, a2, 3 * 4 * 2
+        vle16.v      v2, (t2)
+        vle16.v      v3, (t3)
+        li           t1, 7
+        jal          t0, ff_vc1_inv_trans_4_rvv
+        add          t1, a1, a0
+        vle8.v       v8, (a0)
+        add          t2, a1, t1
+        vle8.v       v9, (t1)
+        add          t3, a1, t2
+        vle8.v       v10, (t2)
+        vle8.v       v11, (t3)
+        vsetvli      zero, zero, e8, mf4, ta, ma
+        vwaddu.wv    v0, v0, v8
+        vwaddu.wv    v1, v1, v9
+        vwaddu.wv    v2, v2, v10
+        vwaddu.wv    v3, v3, v11
+        vsetvli      zero, zero, e16, mf2, ta, ma
+        .irp    n,0,1,2,3
+        vmax.vx      v\n, v\n, zero
+        .endr
+        vsetvli      zero, zero, e8, mf4, ta, ma
+        vnclipu.wi   v8, v0, 0
+        vnclipu.wi   v9, v1, 0
+        vse8.v       v8, (a0)
+        vnclipu.wi   v10, v2, 0
+        vse8.v       v9, (t1)
+        vnclipu.wi   v11, v3, 0
+        vse8.v       v10, (t2)
+        vse8.v       v11, (t3)
+        ret
+endfunc
+
 .macro mspel_op op pos n1 n2
         add           t1, \pos, a2
         v\op\()e8.v   v\n1, (\pos)
-- 
2.45.1