lavu/riscv: count bytes rather than words for bswap32

This removes the dependency on Zba at essentially zero cost.

libavcodec/riscv/bswapdsp_init.c

@@ -33,15 +33,15 @@ av_cold void ff_bswapdsp_init_riscv(BswapDSPContext *c)
 #if HAVE_RV
     int flags = av_get_cpu_flags();
 
-    if (flags & AV_CPU_FLAG_RVB_ADDR) {
 #if (__riscv_xlen >= 64)
     if (flags & AV_CPU_FLAG_RVB_BASIC)
         c->bswap_buf = ff_bswap32_buf_rvb;
 #endif
 #if HAVE_RVV
+    if (flags & AV_CPU_FLAG_RVB_ADDR) {
         if (flags & AV_CPU_FLAG_RVV_I32)
             c->bswap16_buf = ff_bswap16_buf_rvv;
-#endif
     }
 #endif
+#endif
 }

libavcodec/riscv/bswapdsp_rvb.S

@@ -23,8 +23,9 @@
 #include "libavutil/riscv/bswap_rvb.S"
 
 #if (__riscv_xlen >= 64)
-func ff_bswap32_buf_rvb, zba, zbb
+func ff_bswap32_buf_rvb, zbb
         lpad    0
+        slli    a2, a2, 2
         bswap32_rvb a0, a1, a2
 endfunc
 #endif

libavutil/riscv/bswap_rvb.S

@@ -20,21 +20,21 @@
 
 #if (__riscv_xlen >= 64)
         .macro  bswap32_rvb out, in, count
-        andi    t0, \count, 4
+        andi    t0, \count, 1
         beqz    t0, 1f
         /* Align input to 64-bit */
         lwu     t0, (\in)
         addi    \out, \out, 4
         rev8    t0, t0
-        addi    \count, \count, -1
+        addi    \count, \count, -4
         srli    t0, t0, __riscv_xlen - 32
         addi    \in, \in, 4
         sw      t0, -4(\out)
 1:
-        andi    t3, \count, -2
+        andi    t3, \count, -8
-        sh2add  \count, \count, \out
+        add     \count, \count, \out
         beqz    t3, 3f
-        sh2add  t3, t3, \out
+        add     t3, t3, \out
 2:      /* 2 elements (64 bits) at a time on a 64-bit boundary */
         ld      t0,  (\in)
         addi    \out, \out, 8

libswscale/riscv/rgb2rgb.c

@@ -46,7 +46,7 @@ av_cold void rgb2rgb_init_riscv(void)
     int flags = av_get_cpu_flags();
 
 #if (__riscv_xlen == 64)
-    if ((flags & AV_CPU_FLAG_RVB_BASIC) && (flags & AV_CPU_FLAG_RVB_ADDR))
+    if (flags & AV_CPU_FLAG_RVB_BASIC)
         shuffle_bytes_3210 = ff_shuffle_bytes_3210_rvb;
 #endif
 #if HAVE_RVV

libswscale/riscv/rgb2rgb_rvb.S

@@ -23,9 +23,8 @@
 #include "libavutil/riscv/bswap_rvb.S"
 
 #if (__riscv_xlen >= 64)
-func ff_shuffle_bytes_3210_rvb, zba, zbb
+func ff_shuffle_bytes_3210_rvb, zbb
         lpad    0
-        srli    a2, a2, 2
         bswap32_rvb a1, a0, a2
 endfunc
 #endif