lavc/lpc: optimise RVV vector type for compute_autocorr

On SpacemiT X60 (with len == 4000): autocorr_10_c: 2303.7 autocorr_10_rvv_f64: 1411.5 (before) autocorr_10_rvv_f64: 842.2 (after)

Rémi Denis-Courmont committed 2 years ago

Commit c53d4231 parent d82c5032 files · +5 −3Browse files

Unified · Split

2 files

@@ -36,7 +36,8 @@ av_cold void ff_lpc_init_riscv(LPCContext *c) · 35 unmodified lines
36	if ((flags & AV_CPU_FLAG_RVV_F64) && (flags & AV_CPU_FLAG_RVB_ADDR)) {	36	if ((flags & AV_CPU_FLAG_RVV_F64) && (flags & AV_CPU_FLAG_RVB_ADDR)) {
37	c->lpc_apply_welch_window = ff_lpc_apply_welch_window_rvv;	37	c->lpc_apply_welch_window = ff_lpc_apply_welch_window_rvv;
38		38
39	if (ff_get_rv_vlenb() > c->max_order)	39	if ((flags & AV_CPU_FLAG_RVB_BASIC) &&
		40	ff_get_rv_vlenb() > c->max_order)
40	c->lpc_compute_autocorr = ff_lpc_compute_autocorr_rvv;	41	c->lpc_compute_autocorr = ff_lpc_compute_autocorr_rvv;
41	}	42	}
42	#endif	43	#endif

         ret
 endfunc
-func ff_lpc_compute_autocorr_rvv, zve64d
+func ff_lpc_compute_autocorr_rvv, zve64d, zbb
         addi      a2, a2, 1
+        vtype_vli t1, a2, t2, e64, ta, ma
         li        t0, 1
-        vsetvli   zero, a2, e64, m8, ta, ma
+        vsetvl    zero, a2, t1
         fcvt.d.l  ft0, t0
         vle64.v   v0, (a0)
         sh3add    a0, a2, a0   # data += lag

Keyboard shortcuts