From 02594c8c011ca2cdc20334e9bb812ec8f6f37cf3 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?R=C3=A9mi=20Denis-Courmont?= <remi@remlab.net>
Date: Sun, 29 Oct 2023 13:26:30 +0200
Subject: [PATCH] lavc/pixblockdsp: rework R-V V get_pixels_unaligned

As in the aligned case, we can use VLSE64.V, though the way of doing so
gets more convoluted, so the performance gains are more modest:

get_pixels_unaligned_c:       126.7
get_pixels_unaligned_rvv_i32: 145.5 (before)
get_pixels_unaligned_rvv_i64:  62.2 (after)

For the reference, those are the aligned benchmarks (unchanged) on the
same T-Head C908 hardware:

get_pixels_c:                 126.7
get_pixels_rvi:                85.7
get_pixels_rvv_i64:            33.2
---
 libavcodec/riscv/pixblockdsp_init.c | 17 +++++++----------
 libavcodec/riscv/pixblockdsp_rvv.S  | 28 +++++++++++++++++-----------
 2 files changed, 24 insertions(+), 21 deletions(-)

diff --git a/libavcodec/riscv/pixblockdsp_init.c b/libavcodec/riscv/pixblockdsp_init.c
index 6b1efd16f8..3c623a9473 100644
--- a/libavcodec/riscv/pixblockdsp_init.c
+++ b/libavcodec/riscv/pixblockdsp_init.c
@@ -56,20 +56,17 @@ av_cold void ff_pixblockdsp_init_riscv(PixblockDSPContext *c,
 
 #if HAVE_RVV
     if ((cpu_flags & AV_CPU_FLAG_RVV_I32) && ff_get_rv_vlenb() >= 16) {
+        c->diff_pixels = ff_diff_pixels_unaligned_rvv;
+        c->diff_pixels_unaligned = ff_diff_pixels_unaligned_rvv;
+    }
+
+    if ((cpu_flags & AV_CPU_FLAG_RVV_I64) && ff_get_rv_vlenb() >= 16) {
         if (!high_bit_depth) {
-            c->get_pixels = ff_get_pixels_unaligned_8_rvv;
+            c->get_pixels = ff_get_pixels_8_rvv;
             c->get_pixels_unaligned = ff_get_pixels_unaligned_8_rvv;
         }
 
-        c->diff_pixels = ff_diff_pixels_unaligned_rvv;
-        c->diff_pixels_unaligned = ff_diff_pixels_unaligned_rvv;
-
-        if (cpu_flags & AV_CPU_FLAG_RVV_I64) {
-            if (!high_bit_depth)
-                c->get_pixels = ff_get_pixels_8_rvv;
-
-            c->diff_pixels = ff_diff_pixels_rvv;
-        }
+        c->diff_pixels = ff_diff_pixels_rvv;
     }
 #endif
 }
diff --git a/libavcodec/riscv/pixblockdsp_rvv.S b/libavcodec/riscv/pixblockdsp_rvv.S
index 7e35fc5b46..4213cd1b85 100644
--- a/libavcodec/riscv/pixblockdsp_rvv.S
+++ b/libavcodec/riscv/pixblockdsp_rvv.S
@@ -23,6 +23,7 @@
 func ff_get_pixels_8_rvv, zve64x
         vsetivli zero, 8, e8, mf2, ta, ma
         li      t0, 8 * 8
+1:
         vlse64.v v16, (a1), a2
         vsetvli zero, t0, e8, m4, ta, ma
         vwcvtu.x.x.v v8, v16
@@ -30,18 +31,23 @@ func ff_get_pixels_8_rvv, zve64x
         ret
 endfunc
 
-func ff_get_pixels_unaligned_8_rvv, zve32x
-        vsetivli     zero, 8, e8, mf2, ta, ma
-        vlsseg8e8.v  v16, (a1), a2
+func ff_get_pixels_unaligned_8_rvv, zve64x
+        andi    t1, a1, 7
+        vsetivli zero, 8, e64, m4, ta, ma
+        li      t0, 8 * 8
+        beqz    t1, 1b
+        andi    a1, a1, -8
+        slli    t2, t1, 3
+        addi    t1, a1, 8
+        sub     t3, t0, t2
+        vlse64.v v16, (a1), a2
+        vlse64.v v24, (t1), a2
+        vsrl.vx v16, v16, t2
+        vsll.vx v24, v24, t3
+        vor.vv  v16, v16, v24
+        vsetvli zero, t0, e8, m4, ta, ma
         vwcvtu.x.x.v v8, v16
-        vwcvtu.x.x.v v9, v17
-        vwcvtu.x.x.v v10, v18
-        vwcvtu.x.x.v v11, v19
-        vwcvtu.x.x.v v12, v20
-        vwcvtu.x.x.v v13, v21
-        vwcvtu.x.x.v v14, v22
-        vwcvtu.x.x.v v15, v23
-        vsseg8e16.v  v8, (a0)
+        vse16.v v8, (a0)
         ret
 endfunc