mirrors/opencv
2
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2026-09-25 04:09:57 +03:00
Code Issues Packages Projects Releases Wiki Activity

Merge pull request #30000 from Xlawy:opt/rvv-fp32-l2-widening-fma

Browse Source 
core: optimize RVV FP32 L2 norm with widening FMA
This commit is contained in:
Alexander Smorkalov
2026-09-24 10:28:12 +03:00
committed by GitHub
parent 90c7dce3fe 48b2496ca6
commit aa5546b3e4
1 changed files with 3 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files
hal/riscv-rvv/src/core/norm.cpp
+3 -6
View File
@@ -342,8 +342,7 @@ struct NormL2_RVV<float, double> {
for (int i = 0; i < n; i += vl) { for (int i = 0; i < n; i += vl) {
vl = __riscv_vsetvl_e32m4(n - i); vl = __riscv_vsetvl_e32m4(n - i);
auto v = __riscv_vle32_v_f32m4(src + i, vl); auto v = __riscv_vle32_v_f32m4(src + i, vl);
auto v_mul = __riscv_vfwmul(v, v, vl); s = __riscv_vfwmacc_tu(s, v, v, vl);
s = __riscv_vfadd_tu(s, s, v_mul, vl);
} }
return __riscv_vfmv_f(__riscv_vfredosum(s, __riscv_vfmv_s_f_f64m1(0, __riscv_vsetvlmax_e64m1()), vlmax)); return __riscv_vfmv_f(__riscv_vfredosum(s, __riscv_vfmv_s_f_f64m1(0, __riscv_vsetvlmax_e64m1()), vlmax));
} }
@@ -933,8 +932,7 @@ struct MaskedNormL2_RVV<float, double> {
auto v = __riscv_vle32_v_f32m4(src + i, vl); auto v = __riscv_vle32_v_f32m4(src + i, vl);
auto m = __riscv_vle8_v_u8m1(mask + i, vl); auto m = __riscv_vle8_v_u8m1(mask + i, vl);
auto b = __riscv_vmsne(m, 0, vl); auto b = __riscv_vmsne(m, 0, vl);
auto v_mul = __riscv_vfwmul(b, v, v, vl); s = __riscv_vfwmacc_tumu(b, s, v, v, vl);
s = __riscv_vfadd_tumu(b, s, s, v_mul, vl);
} }
} else { } else {
for (int cn_index = 0; cn_index < cn; cn_index++) { for (int cn_index = 0; cn_index < cn; cn_index++) {
@@ -944,8 +942,7 @@ struct MaskedNormL2_RVV<float, double> {
auto v = __riscv_vlse32_v_f32m4(src + cn * i + cn_index, sizeof(float) * cn, vl); auto v = __riscv_vlse32_v_f32m4(src + cn * i + cn_index, sizeof(float) * cn, vl);
auto m = __riscv_vle8_v_u8m1(mask + i, vl); auto m = __riscv_vle8_v_u8m1(mask + i, vl);
auto b = __riscv_vmsne(m, 0, vl); auto b = __riscv_vmsne(m, 0, vl);
auto v_mul = __riscv_vfwmul(b, v, v, vl); s = __riscv_vfwmacc_tumu(b, s, v, v, vl);
s = __riscv_vfadd_tumu(b, s, s, v_mul, vl);
} }
} }
} }