As we do not need to widen accumulators to 64 bits, we effectively get double capacity for unrolling compared to the integer function. This explains the slightly better performance gains. ac3_sum_square_bufferfly_float_c: 65.2 ac3_sum_square_bufferfly_float_rvv_f32: 12.2
Unified · Split
| @@ -30,6 +30,8 @@ void ff_extract_exponents_rvb(uint8_t *exp, int32_t *coef, int nb_coefs); · 29 unmodified lines | ||
| 30 | 30 | void ff_float_to_fixed24_rvv(int32_t *dst, const float *src, size_t len); |
| 31 | 31 | void ff_sum_square_butterfly_int32_rvv(int64_t *, const int32_t *, |
| 32 | 32 | const int32_t *, int); |
| 33 | +void ff_sum_square_butterfly_float_rvv(float *, const float *, | |
| 34 | + const float *, int); | |
| 33 | 35 | |
| 34 | 36 | av_cold void ff_ac3dsp_init_riscv(AC3DSPContext *c) |
| 35 | 37 | { |
| @@ -39,8 +41,10 @@ av_cold void ff_ac3dsp_init_riscv(AC3DSPContext *c) · 3 unmodified lines | ||
| 39 | 41 | if (flags & AV_CPU_FLAG_RVB_ADDR) { |
| 40 | 42 | if (flags & AV_CPU_FLAG_RVB_BASIC) |
| 41 | 43 | c->extract_exponents = ff_extract_exponents_rvb; |
| 42 | − if (flags & AV_CPU_FLAG_RVV_F32) | |
| 44 | + if (flags & AV_CPU_FLAG_RVV_F32) { | |
| 43 | 45 | c->float_to_fixed24 = ff_float_to_fixed24_rvv; |
| 46 | + c->sum_square_butterfly_float = ff_sum_square_butterfly_float_rvv; | |
| 47 | + } | |
| 44 | 48 | # if __riscv_xlen >= 64 |
| 45 | 49 | if (flags & AV_CPU_FLAG_RVV_I64) |
| 46 | 50 | c->sum_square_butterfly_int32 = ff_sum_square_butterfly_int32_rvv; |
| @@ -78,3 +78,42 @@ func ff_sum_square_butterfly_int32_rvv, zve64x · 77 unmodified lines | ||
| 78 | 78 | ret |
| 79 | 79 | endfunc |
| 80 | 80 | #endif |
| 81 | + | |
| 82 | +func ff_sum_square_butterfly_float_rvv, zve32f | |
| 83 | + vsetvli t0, zero, e32, m8, ta, ma | |
| 84 | + vmv.v.x v0, zero | |
| 85 | + vmv.v.x v8, zero | |
| 86 | +1: | |
| 87 | + vsetvli t0, a3, e32, m4, tu, ma | |
| 88 | + vle32.v v16, (a1) | |
| 89 | + sub a3, a3, t0 | |
| 90 | + vle32.v v20, (a2) | |
| 91 | + sh2add a1, t0, a1 | |
| 92 | + vfadd.vv v24, v16, v20 | |
| 93 | + sh2add a2, t0, a2 | |
| 94 | + vfsub.vv v28, v16, v20 | |
| 95 | + vfmacc.vv v0, v16, v16 | |
| 96 | + vfmacc.vv v4, v20, v20 | |
| 97 | + vfmacc.vv v8, v24, v24 | |
| 98 | + vfmacc.vv v12, v28, v28 | |
| 99 | + bnez a3, 1b | |
| 100 | + | |
| 101 | + vsetvli t0, zero, e32, m4, ta, ma | |
| 102 | + vmv.s.x v16, zero | |
| 103 | + vmv.s.x v17, zero | |
| 104 | + vfredsum.vs v16, v0, v16 | |
| 105 | + vmv.s.x v18, zero | |
| 106 | + vfredsum.vs v17, v4, v17 | |
| 107 | + vmv.s.x v19, zero | |
| 108 | + vfredsum.vs v18, v8, v18 | |
| 109 | + vfmv.f.s ft0, v16 | |
| 110 | + vfredsum.vs v19, v12, v19 | |
| 111 | + vfmv.f.s ft1, v17 | |
| 112 | + fsw ft0, (a0) | |
| 113 | + vfmv.f.s ft2, v18 | |
| 114 | + fsw ft1, 4(a0) | |
| 115 | + vfmv.f.s ft3, v19 | |
| 116 | + fsw ft2, 8(a0) | |
| 117 | + fsw ft3, 12(a0) | |
| 118 | + ret | |
| 119 | +endfunc | |