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

rvv: reuse mask chunks across integer norm channels

Browse Source 
Process masked integer norms chunk by chunk and reuse each loaded mask
and predicate across channels, avoiding repeated mask processing.

Extend norm_mask performance coverage with three-channel 8U, 8S, 16U,
16S, and 32S inputs for INF, L1, and L2 norms.

Co-authored-by: Yang Wang <yangwang@iscas.ac.cn>
Co-authored-by: Yuansheng <yuansheng@isrc.iscas.ac.cn>
This commit is contained in:
Zhang Jinhan
2026-09-21 14:18:08 +08:00
co-authored by Yang Wang Yuansheng
parent 8d126c4a67
commit 6c29da7a81
2 changed files with 150 additions and 43 deletions
Download Patch File Download Diff File Expand all files Collapse all files
hal/riscv-rvv/src/core/norm.cpp
+146 -42
View File
@@ -437,13 +437,13 @@ struct MaskedNormInf_RVV<uchar, int> {
s = __riscv_vmaxu_tumu(b, s, s, v, vl * 4);
}
} else {
for (int cn_index = 0; cn_index < cn; cn_index++) {
int vl;
for (int i = 0; i < len; i += vl) {
vl = __riscv_vsetvl_e8m8(len - i);
int vl;
for (int i = 0; i < len; i += vl) {
vl = __riscv_vsetvl_e8m8(len - i);
auto m = __riscv_vle8_v_u8m8(mask + i, vl);
auto b = __riscv_vmsne(m, 0, vl);
for (int cn_index = 0; cn_index < cn; cn_index++) {
auto v = __riscv_vlse8_v_u8m8(src + cn * i + cn_index, sizeof(uchar) * cn, vl);
auto m = __riscv_vle8_v_u8m8(mask + i, vl);
auto b = __riscv_vmsne(m, 0, vl);
s = __riscv_vmaxu_tumu(b, s, s, v, vl);
}
}
@@ -476,13 +476,13 @@ struct MaskedNormL1_RVV<uchar, int> {
s = __riscv_vwredsumu(__riscv_vwredsumu_tum(b, zero, v, zero, vl * 4), s, __riscv_vsetvlmax_e16m1());
}
} else {
for (int cn_index = 0; cn_index < cn; cn_index++) {
int vl;
for (int i = 0; i < len; i += vl) {
vl = __riscv_vsetvl_e8m8(len - i);
int vl;
for (int i = 0; i < len; i += vl) {
vl = __riscv_vsetvl_e8m8(len - i);
auto m = __riscv_vle8_v_u8m8(mask + i, vl);
auto b = __riscv_vmsne(m, 0, vl);
for (int cn_index = 0; cn_index < cn; cn_index++) {
auto v = __riscv_vlse8_v_u8m8(src + cn * i + cn_index, sizeof(uchar) * cn, vl);
auto m = __riscv_vle8_v_u8m8(mask + i, vl);
auto b = __riscv_vmsne(m, 0, vl);
s = __riscv_vwredsumu(__riscv_vwredsumu_tum(b, zero, v, zero, vl), s, __riscv_vsetvlmax_e16m1());
}
}
@@ -514,13 +514,13 @@ struct MaskedNormL2_RVV<uchar, int> {
s = __riscv_vwredsumu(b, __riscv_vwmulu(b, v, v, vl * 4), s, vl * 4);
}
} else {
for (int cn_index = 0; cn_index < cn; cn_index++) {
int vl;
for (int i = 0; i < len; i += vl) {
vl = __riscv_vsetvl_e8m4(len - i);
int vl;
for (int i = 0; i < len; i += vl) {
vl = __riscv_vsetvl_e8m4(len - i);
auto m = __riscv_vle8_v_u8m4(mask + i, vl);
auto b = __riscv_vmsne(m, 0, vl);
for (int cn_index = 0; cn_index < cn; cn_index++) {
auto v = __riscv_vlse8_v_u8m4(src + cn * i + cn_index, sizeof(uchar) * cn, vl);
auto m = __riscv_vle8_v_u8m4(mask + i, vl);
auto b = __riscv_vmsne(m, 0, vl);
s = __riscv_vwredsumu(b, __riscv_vwmulu(b, v, v, vl), s, vl);
}
}
@@ -534,14 +534,27 @@ struct MaskedNormInf_RVV<schar, int> {
int operator() (const schar* src, const uchar* mask, int len, int cn) const {
int vlmax = __riscv_vsetvlmax_e8m8();
auto s = __riscv_vmv_v_x_u8m8(0, vlmax);
for (int cn_index = 0; cn_index < cn; cn_index++) {
if (cn == 1) {
for (int cn_index = 0; cn_index < cn; cn_index++) {
int vl;
for (int i = 0; i < len; i += vl) {
vl = __riscv_vsetvl_e8m8(len - i);
auto v = __riscv_vlse8_v_i8m8(src + cn * i + cn_index, sizeof(schar) * cn, vl);
auto m = __riscv_vle8_v_u8m8(mask + i, vl);
auto b = __riscv_vmsne(m, 0, vl);
s = __riscv_vmaxu_tumu(b, s, s, common::__riscv_vabs(v, vl), vl);
}
}
} else {
int vl;
for (int i = 0; i < len; i += vl) {
vl = __riscv_vsetvl_e8m8(len - i);
auto v = __riscv_vlse8_v_i8m8(src + cn * i + cn_index, sizeof(schar) * cn, vl);
auto m = __riscv_vle8_v_u8m8(mask + i, vl);
auto b = __riscv_vmsne(m, 0, vl);
s = __riscv_vmaxu_tumu(b, s, s, common::__riscv_vabs(v, vl), vl);
for (int cn_index = 0; cn_index < cn; cn_index++) {
auto v = __riscv_vlse8_v_i8m8(src + cn * i + cn_index, sizeof(schar) * cn, vl);
s = __riscv_vmaxu_tumu(b, s, s, common::__riscv_vabs(v, vl), vl);
}
}
}
return __riscv_vmv_x(__riscv_vredmaxu(s, __riscv_vmv_s_x_u8m1(0, __riscv_vsetvlmax_e8m1()), vlmax));
@@ -553,14 +566,27 @@ struct MaskedNormL1_RVV<schar, int> {
int operator() (const schar* src, const uchar* mask, int len, int cn) const {
auto s = __riscv_vmv_v_x_u32m1(0, __riscv_vsetvlmax_e32m1());
auto zero = __riscv_vmv_v_x_u16m1(0, __riscv_vsetvlmax_e16m1());
for (int cn_index = 0; cn_index < cn; cn_index++) {
if (cn == 1) {
for (int cn_index = 0; cn_index < cn; cn_index++) {
int vl;
for (int i = 0; i < len; i += vl) {
vl = __riscv_vsetvl_e8m8(len - i);
auto v = common::__riscv_vabs(__riscv_vlse8_v_i8m8(src + cn * i + cn_index, sizeof(schar) * cn, vl), vl);
auto m = __riscv_vle8_v_u8m8(mask + i, vl);
auto b = __riscv_vmsne(m, 0, vl);
s = __riscv_vwredsumu(__riscv_vwredsumu_tum(b, zero, v, zero, vl), s, __riscv_vsetvlmax_e16m1());
}
}
} else {
int vl;
for (int i = 0; i < len; i += vl) {
vl = __riscv_vsetvl_e8m8(len - i);
auto v = common::__riscv_vabs(__riscv_vlse8_v_i8m8(src + cn * i + cn_index, sizeof(schar) * cn, vl), vl);
auto m = __riscv_vle8_v_u8m8(mask + i, vl);
auto b = __riscv_vmsne(m, 0, vl);
s = __riscv_vwredsumu(__riscv_vwredsumu_tum(b, zero, v, zero, vl), s, __riscv_vsetvlmax_e16m1());
for (int cn_index = 0; cn_index < cn; cn_index++) {
auto v = common::__riscv_vabs(__riscv_vlse8_v_i8m8(src + cn * i + cn_index, sizeof(schar) * cn, vl), vl);
s = __riscv_vwredsumu(__riscv_vwredsumu_tum(b, zero, v, zero, vl), s, __riscv_vsetvlmax_e16m1());
}
}
}
return __riscv_vmv_x(s);
@@ -571,14 +597,27 @@ template<>
struct MaskedNormL2_RVV<schar, int> {
int operator() (const schar* src, const uchar* mask, int len, int cn) const {
auto s = __riscv_vmv_v_x_i32m1(0, __riscv_vsetvlmax_e32m1());
for (int cn_index = 0; cn_index < cn; cn_index++) {
if (cn == 1) {
for (int cn_index = 0; cn_index < cn; cn_index++) {
int vl;
for (int i = 0; i < len; i += vl) {
vl = __riscv_vsetvl_e8m4(len - i);
auto v = __riscv_vlse8_v_i8m4(src + cn * i + cn_index, sizeof(schar) * cn, vl);
auto m = __riscv_vle8_v_u8m4(mask + i, vl);
auto b = __riscv_vmsne(m, 0, vl);
s = __riscv_vwredsum(b, __riscv_vwmul(b, v, v, vl), s, vl);
}
}
} else {
int vl;
for (int i = 0; i < len; i += vl) {
vl = __riscv_vsetvl_e8m4(len - i);
auto v = __riscv_vlse8_v_i8m4(src + cn * i + cn_index, sizeof(schar) * cn, vl);
auto m = __riscv_vle8_v_u8m4(mask + i, vl);
auto b = __riscv_vmsne(m, 0, vl);
s = __riscv_vwredsum(b, __riscv_vwmul(b, v, v, vl), s, vl);
for (int cn_index = 0; cn_index < cn; cn_index++) {
auto v = __riscv_vlse8_v_i8m4(src + cn * i + cn_index, sizeof(schar) * cn, vl);
s = __riscv_vwredsum(b, __riscv_vwmul(b, v, v, vl), s, vl);
}
}
}
return __riscv_vmv_x(s);
@@ -590,14 +629,27 @@ struct MaskedNormInf_RVV<ushort, int> {
int operator() (const ushort* src, const uchar* mask, int len, int cn) const {
int vlmax = __riscv_vsetvlmax_e16m8();
auto s = __riscv_vmv_v_x_u16m8(0, vlmax);
for (int cn_index = 0; cn_index < cn; cn_index++) {
if (cn == 1) {
for (int cn_index = 0; cn_index < cn; cn_index++) {
int vl;
for (int i = 0; i < len; i += vl) {
vl = __riscv_vsetvl_e16m8(len - i);
auto v = __riscv_vlse16_v_u16m8(src + cn * i + cn_index, sizeof(ushort) * cn, vl);
auto m = __riscv_vle8_v_u8m4(mask + i, vl);
auto b = __riscv_vmsne(m, 0, vl);
s = __riscv_vmaxu_tumu(b, s, s, v, vl);
}
}
} else {
int vl;
for (int i = 0; i < len; i += vl) {
vl = __riscv_vsetvl_e16m8(len - i);
auto v = __riscv_vlse16_v_u16m8(src + cn * i + cn_index, sizeof(ushort) * cn, vl);
auto m = __riscv_vle8_v_u8m4(mask + i, vl);
auto b = __riscv_vmsne(m, 0, vl);
s = __riscv_vmaxu_tumu(b, s, s, v, vl);
for (int cn_index = 0; cn_index < cn; cn_index++) {
auto v = __riscv_vlse16_v_u16m8(src + cn * i + cn_index, sizeof(ushort) * cn, vl);
s = __riscv_vmaxu_tumu(b, s, s, v, vl);
}
}
}
return __riscv_vmv_x(__riscv_vredmaxu(s, __riscv_vmv_s_x_u16m1(0, __riscv_vsetvlmax_e16m1()), vlmax));
@@ -608,14 +660,27 @@ template<>
struct MaskedNormL1_RVV<ushort, int> {
int operator() (const ushort* src, const uchar* mask, int len, int cn) const {
auto s = __riscv_vmv_v_x_u32m1(0, __riscv_vsetvlmax_e32m1());
for (int cn_index = 0; cn_index < cn; cn_index++) {
if (cn == 1) {
for (int cn_index = 0; cn_index < cn; cn_index++) {
int vl;
for (int i = 0; i < len; i += vl) {
vl = __riscv_vsetvl_e8m4(len - i);
auto v = __riscv_vlse16_v_u16m8(src + cn * i + cn_index, sizeof(ushort) * cn, vl);
auto m = __riscv_vle8_v_u8m4(mask + i, vl);
auto b = __riscv_vmsne(m, 0, vl);
s = __riscv_vwredsumu_tum(b, s, v, s, vl);
}
}
} else {
int vl;
for (int i = 0; i < len; i += vl) {
vl = __riscv_vsetvl_e8m4(len - i);
auto v = __riscv_vlse16_v_u16m8(src + cn * i + cn_index, sizeof(ushort) * cn, vl);
auto m = __riscv_vle8_v_u8m4(mask + i, vl);
auto b = __riscv_vmsne(m, 0, vl);
s = __riscv_vwredsumu_tum(b, s, v, s, vl);
for (int cn_index = 0; cn_index < cn; cn_index++) {
auto v = __riscv_vlse16_v_u16m8(src + cn * i + cn_index, sizeof(ushort) * cn, vl);
s = __riscv_vwredsumu_tum(b, s, v, s, vl);
}
}
}
return __riscv_vmv_x(s);
@@ -647,14 +712,27 @@ struct MaskedNormInf_RVV<short, int> {
int operator() (const short* src, const uchar* mask, int len, int cn) const {
int vlmax = __riscv_vsetvlmax_e16m8();
auto s = __riscv_vmv_v_x_u16m8(0, vlmax);
for (int cn_index = 0; cn_index < cn; cn_index++) {
if (cn == 1) {
for (int cn_index = 0; cn_index < cn; cn_index++) {
int vl;
for (int i = 0; i < len; i += vl) {
vl = __riscv_vsetvl_e16m8(len - i);
auto v = __riscv_vlse16_v_i16m8(src + cn * i + cn_index, sizeof(short) * cn, vl);
auto m = __riscv_vle8_v_u8m4(mask + i, vl);
auto b = __riscv_vmsne(m, 0, vl);
s = __riscv_vmaxu_tumu(b, s, s, common::__riscv_vabs(v, vl), vl);
}
}
} else {
int vl;
for (int i = 0; i < len; i += vl) {
vl = __riscv_vsetvl_e16m8(len - i);
auto v = __riscv_vlse16_v_i16m8(src + cn * i + cn_index, sizeof(short) * cn, vl);
auto m = __riscv_vle8_v_u8m4(mask + i, vl);
auto b = __riscv_vmsne(m, 0, vl);
s = __riscv_vmaxu_tumu(b, s, s, common::__riscv_vabs(v, vl), vl);
for (int cn_index = 0; cn_index < cn; cn_index++) {
auto v = __riscv_vlse16_v_i16m8(src + cn * i + cn_index, sizeof(short) * cn, vl);
s = __riscv_vmaxu_tumu(b, s, s, common::__riscv_vabs(v, vl), vl);
}
}
}
return __riscv_vmv_x(__riscv_vredmaxu(s, __riscv_vmv_s_x_u16m1(0, __riscv_vsetvlmax_e16m1()), vlmax));
@@ -665,14 +743,27 @@ template<>
struct MaskedNormL1_RVV<short, int> {
int operator() (const short* src, const uchar* mask, int len, int cn) const {
auto s = __riscv_vmv_v_x_u32m1(0, __riscv_vsetvlmax_e32m1());
for (int cn_index = 0; cn_index < cn; cn_index++) {
if (cn == 1) {
for (int cn_index = 0; cn_index < cn; cn_index++) {
int vl;
for (int i = 0; i < len; i += vl) {
vl = __riscv_vsetvl_e8m4(len - i);
auto v = common::__riscv_vabs(__riscv_vlse16_v_i16m8(src + cn * i + cn_index, sizeof(short) * cn, vl), vl);
auto m = __riscv_vle8_v_u8m4(mask + i, vl);
auto b = __riscv_vmsne(m, 0, vl);
s = __riscv_vwredsumu_tum(b, s, v, s, vl);
}
}
} else {
int vl;
for (int i = 0; i < len; i += vl) {
vl = __riscv_vsetvl_e8m4(len - i);
auto v = common::__riscv_vabs(__riscv_vlse16_v_i16m8(src + cn * i + cn_index, sizeof(short) * cn, vl), vl);
auto m = __riscv_vle8_v_u8m4(mask + i, vl);
auto b = __riscv_vmsne(m, 0, vl);
s = __riscv_vwredsumu_tum(b, s, v, s, vl);
for (int cn_index = 0; cn_index < cn; cn_index++) {
auto v = common::__riscv_vabs(__riscv_vlse16_v_i16m8(src + cn * i + cn_index, sizeof(short) * cn, vl), vl);
s = __riscv_vwredsumu_tum(b, s, v, s, vl);
}
}
}
return __riscv_vmv_x(s);
@@ -704,14 +795,27 @@ struct MaskedNormInf_RVV<int, int> {
int operator() (const int* src, const uchar* mask, int len, int cn) const {
int vlmax = __riscv_vsetvlmax_e32m8();
auto s = __riscv_vmv_v_x_u32m8(0, vlmax);
for (int cn_index = 0; cn_index < cn; cn_index++) {
if (cn == 1) {
for (int cn_index = 0; cn_index < cn; cn_index++) {
int vl;
for (int i = 0; i < len; i += vl) {
vl = __riscv_vsetvl_e32m8(len - i);
auto v = __riscv_vlse32_v_i32m8(src + cn * i + cn_index, sizeof(int) * cn, vl);
auto m = __riscv_vle8_v_u8m2(mask + i, vl);
auto b = __riscv_vmsne(m, 0, vl);
s = __riscv_vmaxu_tumu(b, s, s, common::__riscv_vabs(v, vl), vl);
}
}
} else {
int vl;
for (int i = 0; i < len; i += vl) {
vl = __riscv_vsetvl_e32m8(len - i);
auto v = __riscv_vlse32_v_i32m8(src + cn * i + cn_index, sizeof(int) * cn, vl);
auto m = __riscv_vle8_v_u8m2(mask + i, vl);
auto b = __riscv_vmsne(m, 0, vl);
s = __riscv_vmaxu_tumu(b, s, s, common::__riscv_vabs(v, vl), vl);
for (int cn_index = 0; cn_index < cn; cn_index++) {
auto v = __riscv_vlse32_v_i32m8(src + cn * i + cn_index, sizeof(int) * cn, vl);
s = __riscv_vmaxu_tumu(b, s, s, common::__riscv_vabs(v, vl), vl);
}
}
}
return __riscv_vmv_x(__riscv_vredmaxu(s, __riscv_vmv_s_x_u32m1(0, __riscv_vsetvlmax_e32m1()), vlmax));
modules/core/perf/perf_norm.cpp
+4 -1
View File
@@ -36,7 +36,10 @@ PERF_TEST_P(Size_MatType_NormType, norm,
PERF_TEST_P(Size_MatType_NormType, norm_mask,
testing::Combine(
testing::Values(TYPICAL_MAT_SIZES),
testing::Values(CV_8UC1, CV_8UC4, CV_8SC1, CV_16UC1, CV_16SC1, CV_32SC1, CV_32FC1, CV_64FC1),
testing::Values(CV_8UC1, CV_8UC3, CV_8UC4,
CV_8SC1, CV_8SC3, CV_16UC1, CV_16UC3,
CV_16SC1, CV_16SC3, CV_32SC1, CV_32SC3,
CV_32FC1, CV_64FC1),
testing::Values((int)NORM_INF, (int)NORM_L1, (int)NORM_L2)
)
)