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
Unified · Split
| @@ -56,20 +56,17 @@ av_cold void ff_pixblockdsp_init_riscv(PixblockDSPContext *c, · 55 unmodified lines | |||
| 56 | 56 | ||
| 57 | #if HAVE_RVV | 57 | #if HAVE_RVV |
| 58 | if ((cpu_flags & AV_CPU_FLAG_RVV_I32) && ff_get_rv_vlenb() >= 16) { | 58 | if ((cpu_flags & AV_CPU_FLAG_RVV_I32) && ff_get_rv_vlenb() >= 16) { |
| 59 | if (!high_bit_depth) { | ||
| 60 | c->get_pixels = ff_get_pixels_unaligned_8_rvv; | ||
| 61 | c->get_pixels_unaligned = ff_get_pixels_unaligned_8_rvv; | ||
| 62 | } | ||
| 63 | |||
| 64 | c->diff_pixels = ff_diff_pixels_unaligned_rvv; | 59 | c->diff_pixels = ff_diff_pixels_unaligned_rvv; |
| 65 | c->diff_pixels_unaligned = ff_diff_pixels_unaligned_rvv; | 60 | c->diff_pixels_unaligned = ff_diff_pixels_unaligned_rvv; |
| 61 | } | ||
| 66 | 62 | ||
| 67 | if (cpu_flags & AV_CPU_FLAG_RVV_I64) { | 63 | if ((cpu_flags & AV_CPU_FLAG_RVV_I64) && ff_get_rv_vlenb() >= 16) { |
| 68 | if (!high_bit_depth) | 64 | if (!high_bit_depth) { |
| 69 | c->get_pixels = ff_get_pixels_8_rvv; | 65 | c->get_pixels = ff_get_pixels_8_rvv; |
| 70 | 66 | c->get_pixels_unaligned = ff_get_pixels_unaligned_8_rvv; | |
| 71 | c->diff_pixels = ff_diff_pixels_rvv; | ||
| 72 | } | 67 | } |
| 68 | |||
| 69 | c->diff_pixels = ff_diff_pixels_rvv; | ||
| 73 | } | 70 | } |
| 74 | #endif | 71 | #endif |
| 75 | } | 72 | } |
| @@ -23,6 +23,7 @@ · 22 unmodified lines | |||
| 23 | func ff_get_pixels_8_rvv, zve64x | 23 | func ff_get_pixels_8_rvv, zve64x |
| 24 | vsetivli zero, 8, e8, mf2, ta, ma | 24 | vsetivli zero, 8, e8, mf2, ta, ma |
| 25 | li t0, 8 * 8 | 25 | li t0, 8 * 8 |
| 26 | 1: | ||
| 26 | vlse64.v v16, (a1), a2 | 27 | vlse64.v v16, (a1), a2 |
| 27 | vsetvli zero, t0, e8, m4, ta, ma | 28 | vsetvli zero, t0, e8, m4, ta, ma |
| 28 | vwcvtu.x.x.v v8, v16 | 29 | vwcvtu.x.x.v v8, v16 |
| @@ -30,18 +31,23 @@ func ff_get_pixels_8_rvv, zve64x · 1 unmodified line | |||
| 30 | ret | 31 | ret |
| 31 | endfunc | 32 | endfunc |
| 32 | 33 | ||
| 33 | func ff_get_pixels_unaligned_8_rvv, zve32x | 34 | func ff_get_pixels_unaligned_8_rvv, zve64x |
| 34 | vsetivli zero, 8, e8, mf2, ta, ma | 35 | andi t1, a1, 7 |
| 35 | vlsseg8e8.v v16, (a1), a2 | 36 | vsetivli zero, 8, e64, m4, ta, ma |
| 37 | li t0, 8 * 8 | ||
| 38 | beqz t1, 1b | ||
| 39 | andi a1, a1, -8 | ||
| 40 | slli t2, t1, 3 | ||
| 41 | addi t1, a1, 8 | ||
| 42 | sub t3, t0, t2 | ||
| 43 | vlse64.v v16, (a1), a2 | ||
| 44 | vlse64.v v24, (t1), a2 | ||
| 45 | vsrl.vx v16, v16, t2 | ||
| 46 | vsll.vx v24, v24, t3 | ||
| 47 | vor.vv v16, v16, v24 | ||
| 48 | vsetvli zero, t0, e8, m4, ta, ma | ||
| 36 | vwcvtu.x.x.v v8, v16 | 49 | vwcvtu.x.x.v v8, v16 |
| 37 | vwcvtu.x.x.v v9, v17 | 50 | vse16.v v8, (a0) |
| 38 | vwcvtu.x.x.v v10, v18 | ||
| 39 | vwcvtu.x.x.v v11, v19 | ||
| 40 | vwcvtu.x.x.v v12, v20 | ||
| 41 | vwcvtu.x.x.v v13, v21 | ||
| 42 | vwcvtu.x.x.v v14, v22 | ||
| 43 | vwcvtu.x.x.v v15, v23 | ||
| 44 | vsseg8e16.v v8, (a0) | ||
| 45 | ret | 51 | ret |
| 46 | endfunc | 52 | endfunc |
| 47 | 53 | ||