[FFmpeg-devel] [PATCH 4/5] lavc/h264dsp: reuse the R-V V IDCT DC add functions

[Rémi Denis-Courmont]

This reuses the DC bypass functions from the multiple IDCT functions, to
leverage vector code.

As an added bonus, the caller functions can now rely on the callee functions
to preserve their parameters, thus cutting down on stack spills.
---
 libavcodec/riscv/h264idct_rvv.S | 76 +++++++--------------------------
 1 file changed, 16 insertions(+), 60 deletions(-)

diff --git a/libavcodec/riscv/h264idct_rvv.S b/libavcodec/riscv/h264idct_rvv.S
index 2648e06aeb..c42db6ef29 100644
--- a/libavcodec/riscv/h264idct_rvv.S
+++ b/libavcodec/riscv/h264idct_rvv.S
@@ -536,7 +536,7 @@ endconst
 .macro  idct4_adds type, depth
 func ff_h264_idct_add\type\()_\depth\()_rvv, zve32x
         csrwi   vxrm, 0
-        addi    sp, sp, -96
+        addi    sp, sp, -64
         lla     t0, ff_h264_scan8
         sd      s0,   (sp)
         li      t1, 32 * (\depth / 8)
@@ -547,14 +547,6 @@ func ff_h264_idct_add\type\()_\depth\()_rvv, zve32x
         sd      s3, 32(sp)
         sd      s4, 40(sp)
         sd      s5, 48(sp)
-        sd      s6, 56(sp)
-        sd      s7, 64(sp)
-.if \depth > 8
-        sd      s8, 72(sp)
-        sd      s9, 80(sp)
-        mv      s8, a5
-        mv      s9, a6
-.endif
         vsetivli  zero, 16, e8, m1, ta, ma
         vle8.v    v8, (t0)
 .if \depth == 8
@@ -583,8 +575,8 @@ func ff_h264_idct_add\type\()_\depth\()_rvv, zve32x
         li      s1, 16
         mv      s4, a0
         mv      s5, a1
-        mv      s6, a2
-        mv      s7, a3
+        mv      a1, a2
+        mv      a2, a3
 1:
         andi    t0, s2, 1
         addi    s1, s1, -1
@@ -594,12 +586,7 @@ func ff_h264_idct_add\type\()_\depth\()_rvv, zve32x
 .endif
         lw      t2, (s5)   # block_offset[i]
         andi    t1, s3, 1
-        mv      a1, s6
-        mv      a2, s7
         add     a0, s4, t2
-.if \depth > 8
-        mv      a5, s8
-.endif
 .ifc \type, 16
         bnez    t1, 2f     # if (nnz == 1 && block[i * 16])
 .else
@@ -611,23 +598,13 @@ func ff_h264_idct_add\type\()_\depth\()_rvv, zve32x
 .ifnc \type, 16
         beqz    t1, 3f    # if (block[i * 16])
 .endif
-.if \depth == 8
-        call    ff_h264_idct_dc_add_\depth\()_c
-.else
-        jalr    s9
-.endif
+        jal     ff_h264_idct4_dc_add_\depth\()_rvv
 3:
         srli    s3, s3, 1
         addi    s5, s5, 4
-        addi    s6, s6, 16 * 2 * (\depth / 8)
+        addi    a1, a1, 16 * 2 * (\depth / 8)
         bnez    s1, 1b
 
-.if \depth > 8
-        ld      s9, 80(sp)
-        ld      s8, 72(sp)
-.endif
-        ld      s7, 64(sp)
-        ld      s6, 56(sp)
         ld      s5, 48(sp)
         ld      s4, 40(sp)
         ld      s3, 32(sp)
@@ -635,7 +612,7 @@ func ff_h264_idct_add\type\()_\depth\()_rvv, zve32x
         ld      s1, 16(sp)
         ld      ra,  8(sp)
         ld      s0,  0(sp)
-        addi    sp, sp, 96
+        addi    sp, sp, 64
         ret
 endfunc
 .endm
@@ -646,7 +623,7 @@ idct4_adds 16intra, \depth
 
 func ff_h264_idct8_add4_\depth\()_rvv, zve32x
         csrwi       vxrm, 0
-        addi    sp, sp, -96
+        addi    sp, sp, -64
         lla     t0, ff_h264_scan8
         sd      s0,   (sp)
         li      t1, 4 * 32 * (\depth / 8)
@@ -658,14 +635,6 @@ func ff_h264_idct8_add4_\depth\()_rvv, zve32x
         sd      s3, 32(sp)
         sd      s4, 40(sp)
         sd      s5, 48(sp)
-        sd      s6, 56(sp)
-        sd      s7, 64(sp)
-.if \depth > 8
-        sd      s8, 72(sp)
-        sd      s9, 80(sp)
-        mv      s8, a5
-        mv      s9, a6
-.endif
         vsetivli  zero, 4, e8, mf4, ta, ma
         vlse8.v   v8, (t0), t2
 .if \depth == 8
@@ -689,8 +658,8 @@ func ff_h264_idct8_add4_\depth\()_rvv, zve32x
         li      s1, 4
         mv      s4, a0
         mv      s5, a1
-        mv      s6, a2
-        mv      s7, a3
+        mv      a1, a2
+        mv      a2, a3
 1:
         andi    t0, s2, 1
         addi    s1, s1, -1
@@ -698,33 +667,23 @@ func ff_h264_idct8_add4_\depth\()_rvv, zve32x
         beqz    t0, 3f     # if (nnz)
         lw      t2, (s5)   # block_offset[i]
         andi    t1, s3, 1
-        mv      a1, s6
-        mv      a2, s7
         add     a0, s4, t2
-.if \depth > 8
-        mv      a5, s8
-.endif
         bnez    t1, 2f    # if (nnz == 1 && block[i * 16])
         jal     .Lidct8_add_\depth\()_rvv
-        j       3f
-2:
 .if \depth == 8
-        call    ff_h264_idct8_dc_add_\depth\()_c
+        j       3f
 .else
-        jalr    s9
+        j       4f        # idct8_add_16 updates a1
 .endif
+2:
+        jal     ff_h264_idct8_dc_add_\depth\()_rvv
 3:
+        addi    a1, a1, 4 * 16 * 2 * (\depth / 8)
+4:
         srli    s3, s3, 1
         addi    s5, s5, 4 * 4
-        addi    s6, s6, 4 * 16 * 2 * (\depth / 8)
         bnez    s1, 1b
 
-.if \depth > 8
-        ld      s9, 80(sp)
-        ld      s8, 72(sp)
-.endif
-        ld      s7, 64(sp)
-        ld      s6, 56(sp)
         ld      s5, 48(sp)
         ld      s4, 40(sp)
         ld      s3, 32(sp)
@@ -732,7 +691,7 @@ func ff_h264_idct8_add4_\depth\()_rvv, zve32x
         ld      s1, 16(sp)
         ld      ra,  8(sp)
         ld      s0,  0(sp)
-        addi    sp, sp, 96
+        addi    sp, sp, 64
         ret
 endfunc
 .endr
@@ -740,19 +699,16 @@ endfunc
 .irp    depth, 9, 10, 12, 14
 func ff_h264_idct_add16_\depth\()_rvv, zve32x
         li      a5, (1 << \depth) - 1
-        lla     a6, ff_h264_idct_dc_add_\depth\()_c
         j       ff_h264_idct_add16_16_rvv
 endfunc
 
 func ff_h264_idct_add16intra_\depth\()_rvv, zve32x
         li      a5, (1 << \depth) - 1
-        lla     a6, ff_h264_idct_dc_add_\depth\()_c
         j       ff_h264_idct_add16intra_16_rvv
 endfunc
 
 func ff_h264_idct8_add4_\depth\()_rvv, zve32x
         li      a5, (1 << \depth) - 1
-        lla     a6, ff_h264_idct8_dc_add_\depth\()_c
         j       ff_h264_idct8_add4_16_rvv
 endfunc
 .endr
-- 
2.45.2