diff options
| author | Chih-Min Chao <chihmin.chao@sifive.com> | 2019-05-15 23:48:02 -0700 |
|---|---|---|
| committer | Chih-Min Chao <chihmin.chao@sifive.com> | 2019-05-19 21:26:10 -0700 |
| commit | ce94de9459f5a1b88c2827ebe75c6589831b474b (patch) | |
| tree | 8f99817dd06c1ec5c80295bb22ff484ab6195322 | |
| parent | 9641f5487569a204267c34d1037e4f3dceba54ff (diff) | |
| download | spike-ce94de9459f5a1b88c2827ebe75c6589831b474b.zip spike-ce94de9459f5a1b88c2827ebe75c6589831b474b.tar.gz spike-ce94de9459f5a1b88c2827ebe75c6589831b474b.tar.bz2 | |
rvv: add tail zeroing for floating reduction
Signed-off-by: Chih-Min Chao <chihmin.chao@sifive.com>
| -rw-r--r-- | riscv/decode.h | 17 |
1 files changed, 13 insertions, 4 deletions
diff --git a/riscv/decode.h b/riscv/decode.h index 7071abb..2e472e9 100644 --- a/riscv/decode.h +++ b/riscv/decode.h @@ -1217,6 +1217,17 @@ VI_LOOP_END P.VU.vstart = 0; \ set_fp_exceptions; +#define VF_LOOP_REDUCTION_END(x) \ + } \ + P.VU.vstart = 0; \ + set_fp_exceptions; \ + if (vl > 0) { \ + P.VU.elt<type_sew_t<x>::type>(rd_num, 0) = vd_0.v; \ + for (reg_t i = 1; i < P.VU.VLEN / x; ++i) { \ + P.VU.elt<type_sew_t<x>::type>(rd_num, i) = 0; \ + } \ + } + #define VF_LOOP_CMP_END \ switch(P.VU.vsew) { \ case 32: { \ @@ -1266,16 +1277,14 @@ VI_LOOP_END BODY; \ vd = i == 0? vd: 0;\ DEBUG_RVV_FP_VV; \ - VF_LOOP_END \ - P.VU.elt<float32_t>(rd_num, 0) = vd_0; + VF_LOOP_REDUCTION_END(e32) #define VFP_VV_LOOP_WIDE_REDUCTION(BODY) \ VF_LOOP_WIDE_REDUCTION_BASE \ float64_t vs2 = f64(P.VU.elt<float32_t>(rs2_num, i).v); \ BODY; \ DEBUG_RVV_FP_VV; \ - VF_LOOP_END \ - P.VU.elt<float64_t>(rd_num, 0) = vd_0; + VF_LOOP_REDUCTION_END(e64) #define VFP_VF_LOOP(BODY) \ VF_LOOP_BASE \ |