mirror of https://git.ffmpeg.org/ffmpeg.git
lavu/fixed_dsp: optimise R-V V fmul_reverse
Gathers are (unsurprisingly) a notable exception to the rule that R-V V gets faster with larger group multipliers. So roll the function to speed it up. Before: vector_fmul_reverse_fixed_c: 2840.7 vector_fmul_reverse_fixed_rvv_i32: 2430.2 After: vector_fmul_reverse_fixed_c: 2841.0 vector_fmul_reverse_fixed_rvv_i32: 962.2 It might be possible to further optimise the function by moving the reverse-subtract out of the loop and adding ad-hoc tail handling.
This commit is contained in:
parent
4adb93dff0
commit
3a134e8299
|
@ -83,16 +83,17 @@ endfunc
|
||||||
|
|
||||||
func ff_vector_fmul_reverse_fixed_rvv, zve32x
|
func ff_vector_fmul_reverse_fixed_rvv, zve32x
|
||||||
csrwi vxrm, 0
|
csrwi vxrm, 0
|
||||||
vsetvli t0, zero, e16, m4, ta, ma
|
// e16/m4 and e32/m8 are possible but slow the gathers down.
|
||||||
|
vsetvli t0, zero, e16, m1, ta, ma
|
||||||
sh2add a2, a3, a2
|
sh2add a2, a3, a2
|
||||||
vid.v v0
|
vid.v v0
|
||||||
vadd.vi v0, v0, 1
|
vadd.vi v0, v0, 1
|
||||||
1:
|
1:
|
||||||
vsetvli t0, a3, e16, m4, ta, ma
|
vsetvli t0, a3, e16, m1, ta, ma
|
||||||
slli t1, t0, 2
|
slli t1, t0, 2
|
||||||
vrsub.vx v4, v0, t0 // v4[i] = [VL-1, VL-2... 1, 0]
|
vrsub.vx v4, v0, t0 // v4[i] = [VL-1, VL-2... 1, 0]
|
||||||
sub a2, a2, t1
|
sub a2, a2, t1
|
||||||
vsetvli zero, zero, e32, m8, ta, ma
|
vsetvli zero, zero, e32, m2, ta, ma
|
||||||
vle32.v v8, (a2)
|
vle32.v v8, (a2)
|
||||||
sub a3, a3, t0
|
sub a3, a3, t0
|
||||||
vle32.v v16, (a1)
|
vle32.v v16, (a1)
|
||||||
|
|
Loading…
Reference in New Issue