Segmented loads may be slower than not. So this advantageously uses a unit-strided load and narrowing shifts instead. Before: ps_add_squares_c: 60757.7 ps_add_squares_rvv_f32: 22242.5 After: ps_add_squares_c: 60516.0 ps_add_squares_rvv_i64: 17067.7
Unified · Split
| @@ -46,7 +46,8 @@ av_cold void ff_psdsp_init_riscv(PSDSPContext *c) · 45 unmodified lines | ||
| 46 | 46 | c->hybrid_analysis = ff_ps_hybrid_analysis_rvv; |
| 47 | 47 | |
| 48 | 48 | if (flags & AV_CPU_FLAG_RVB_ADDR) { |
| 49 | − c->add_squares = ff_ps_add_squares_rvv; | |
| 49 | + if (flags & AV_CPU_FLAG_RVV_I64) | |
| 50 | + c->add_squares = ff_ps_add_squares_rvv; | |
| 50 | 51 | c->mul_pair_single = ff_ps_mul_pair_single_rvv; |
| 51 | 52 | c->stereo_interpolate[0] = ff_ps_stereo_interpolate_rvv; |
| 52 | 53 | } |
| @@ -1,5 +1,5 @@ | ||
| 1 | 1 | /* |
| 2 | − * Copyright © 2022 Rémi Denis-Courmont. | |
| 2 | + * Copyright © 2022-2023 Rémi Denis-Courmont. | |
| 3 | 3 | * |
| 4 | 4 | * This file is part of FFmpeg. |
| 5 | 5 | * |
| @@ -20,13 +20,16 @@ · 14 unmodified lines | ||
| 20 | 20 | |
| 21 | 21 | #include "libavutil/riscv/asm.S" |
| 22 | 22 | |
| 23 | −func ff_ps_add_squares_rvv, zve32f | |
| 23 | +func ff_ps_add_squares_rvv, zve64f | |
| 24 | + li t1, 32 | |
| 24 | 25 | 1: |
| 25 | 26 | vsetvli t0, a2, e32, m4, ta, ma |
| 26 | − vlseg2e32.v v24, (a1) | |
| 27 | + vle64.v v8, (a1) | |
| 27 | 28 | sub a2, a2, t0 |
| 29 | + vnsrl.wx v24, v8, zero | |
| 28 | 30 | vle32.v v16, (a0) |
| 29 | 31 | sh3add a1, t0, a1 |
| 32 | + vnsrl.wx v28, v8, t1 | |
| 30 | 33 | vfmacc.vv v16, v24, v24 |
| 31 | 34 | vfmacc.vv v16, v28, v28 |
| 32 | 35 | vse32.v v16, (a0) |