mirror/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2025-08-06 06:26:29 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

Add test cases and fix bugs in the RVV HAL.

Browse Source
This commit is contained in:
Liutong HAN 2024-12-27 08:39:52 +00:00
parent 94bccbecc0
commit b31f7694c5
3 changed files with 83 additions and 37 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
3rdparty/hal_rvv/hal_rvv_1p0/mean.hpp vendored
View File

@ -119,8 +119,8 @@ inline int meanStdDev_8UC4(const uchar* src_data, size_t src_step, int width, in
vec_sqsum = __riscv_vwmaccu_vv_u64m8_tumu(vmask, vec_sqsum, vec_pixel, vec_pixel, vl);
nz += __riscv_vcpop_m_b8(vmask, vl);
}
nz /= 4;
}
nz /= 4;
} else {
for (int i = 0; i < height; i++) {
const uchar* src_row = src_data + i * src_step;

95
3rdparty/hal_rvv/hal_rvv_1p0/merge.hpp vendored
View File

@ -20,9 +20,9 @@ namespace cv { namespace cv_hal_rvv {
#if defined __GNUC__
__attribute__((optimize("no-tree-vectorize")))
#endif
static int merge8u(const uchar** src, uchar* dst, int len, int cn ) {
inline int merge8u(const uchar** src, uchar* dst, int len, int cn ) {
int k = cn % 4 ? cn % 4 : 4;
int i = 0, j;
int i = 0;
int vl = __riscv_vsetvlmax_e8m1();
if( k == 1 )
{
@ -30,7 +30,7 @@ static int merge8u(const uchar** src, uchar* dst, int len, int cn ) {
for( ; i <= len - vl; i += vl)
{
auto a = __riscv_vle8_v_u8m1(src0 + i, vl);
__riscv_vsse8_v_u8m1(dst + i*cn, sizeof(uchar)*2, a, vl);
__riscv_vsse8_v_u8m1(dst + i*cn, sizeof(uchar)*cn, a, vl);
}
#if defined(__clang__)
#pragma clang loop vectorize(disable)
@ -45,8 +45,8 @@ static int merge8u(const uchar** src, uchar* dst, int len, int cn ) {
{
auto a = __riscv_vle8_v_u8m1(src0 + i, vl);
auto b = __riscv_vle8_v_u8m1(src1 + i, vl);
__riscv_vsse8_v_u8m1(dst + i*cn, sizeof(uchar)*2, a, vl);
__riscv_vsse8_v_u8m1(dst + i*cn + 1, sizeof(uchar)*2, b, vl);
__riscv_vsse8_v_u8m1(dst + i*cn, sizeof(uchar)*cn, a, vl);
__riscv_vsse8_v_u8m1(dst + i*cn + 1, sizeof(uchar)*cn, b, vl);
}
#if defined(__clang__)
#pragma clang loop vectorize(disable)
@ -65,9 +65,9 @@ static int merge8u(const uchar** src, uchar* dst, int len, int cn ) {
auto a = __riscv_vle8_v_u8m1(src0 + i, vl);
auto b = __riscv_vle8_v_u8m1(src1 + i, vl);
auto c = __riscv_vle8_v_u8m1(src2 + i, vl);
__riscv_vsse8_v_u8m1(dst + i*cn, sizeof(uchar)*3, a, vl);
__riscv_vsse8_v_u8m1(dst + i*cn + 1, sizeof(uchar)*3, b, vl);
__riscv_vsse8_v_u8m1(dst + i*cn + 2, sizeof(uchar)*3, c, vl);
__riscv_vsse8_v_u8m1(dst + i*cn, sizeof(uchar)*cn, a, vl);
__riscv_vsse8_v_u8m1(dst + i*cn + 1, sizeof(uchar)*cn, b, vl);
__riscv_vsse8_v_u8m1(dst + i*cn + 2, sizeof(uchar)*cn, c, vl);
}
#if defined(__clang__)
#pragma clang loop vectorize(disable)
@ -88,10 +88,10 @@ static int merge8u(const uchar** src, uchar* dst, int len, int cn ) {
auto b = __riscv_vle8_v_u8m1(src1 + i, vl);
auto c = __riscv_vle8_v_u8m1(src2 + i, vl);
auto d = __riscv_vle8_v_u8m1(src3 + i, vl);
__riscv_vsse8_v_u8m1(dst + i*cn, sizeof(uchar)*4, a, vl);
__riscv_vsse8_v_u8m1(dst + i*cn + 1, sizeof(uchar)*4, b, vl);
__riscv_vsse8_v_u8m1(dst + i*cn + 2, sizeof(uchar)*4, c, vl);
__riscv_vsse8_v_u8m1(dst + i*cn + 3, sizeof(uchar)*4, d, vl);
__riscv_vsse8_v_u8m1(dst + i*cn, sizeof(uchar)*cn, a, vl);
__riscv_vsse8_v_u8m1(dst + i*cn + 1, sizeof(uchar)*cn, b, vl);
__riscv_vsse8_v_u8m1(dst + i*cn + 2, sizeof(uchar)*cn, c, vl);
__riscv_vsse8_v_u8m1(dst + i*cn + 3, sizeof(uchar)*cn, d, vl);
}
#if defined(__clang__)
#pragma clang loop vectorize(disable)
@ -110,10 +110,27 @@ static int merge8u(const uchar** src, uchar* dst, int len, int cn ) {
for( ; k < cn; k += 4 )
{
const uchar *src0 = src[k], *src1 = src[k+1], *src2 = src[k+2], *src3 = src[k+3];
for( i = 0, j = k; i < len; i++, j += cn )
i = 0;
for( ; i <= len - vl; i += vl)
{
dst[j] = src0[i]; dst[j+1] = src1[i];
dst[j+2] = src2[i]; dst[j+3] = src3[i];
auto a = __riscv_vle8_v_u8m1(src0 + i, vl);
auto b = __riscv_vle8_v_u8m1(src1 + i, vl);
auto c = __riscv_vle8_v_u8m1(src2 + i, vl);
auto d = __riscv_vle8_v_u8m1(src3 + i, vl);
__riscv_vsse8_v_u8m1(dst + k+i*cn, sizeof(uchar)*cn, a, vl);
__riscv_vsse8_v_u8m1(dst + k+i*cn + 1, sizeof(uchar)*cn, b, vl);
__riscv_vsse8_v_u8m1(dst + k+i*cn + 2, sizeof(uchar)*cn, c, vl);
__riscv_vsse8_v_u8m1(dst + k+i*cn + 3, sizeof(uchar)*cn, d, vl);
}
#if defined(__clang__)
#pragma clang loop vectorize(disable)
#endif
for( ; i < len; i++ )
{
dst[k+i*cn] = src0[i];
dst[k+i*cn+1] = src1[i];
dst[k+i*cn+2] = src2[i];
dst[k+i*cn+3] = src3[i];
}
}
return CV_HAL_ERROR_OK;
@ -122,9 +139,9 @@ static int merge8u(const uchar** src, uchar* dst, int len, int cn ) {
#if defined __GNUC__
__attribute__((optimize("no-tree-vectorize")))
#endif
static int merge16u(const ushort** src, ushort* dst, int len, int cn ) {
inline int merge16u(const ushort** src, ushort* dst, int len, int cn ) {
int k = cn % 4 ? cn % 4 : 4;
int i = 0, j;
int i = 0;
int vl = __riscv_vsetvlmax_e16m1();
if( k == 1 )
{
@ -132,7 +149,7 @@ static int merge16u(const ushort** src, ushort* dst, int len, int cn ) {
for( ; i <= len - vl; i += vl)
{
auto a = __riscv_vle16_v_u16m1(src0 + i, vl);
__riscv_vsse16_v_u16m1(dst + i*cn, sizeof(ushort)*2, a, vl);
__riscv_vsse16_v_u16m1(dst + i*cn, sizeof(ushort)*cn, a, vl);
}
#if defined(__clang__)
#pragma clang loop vectorize(disable)
@ -147,8 +164,8 @@ static int merge16u(const ushort** src, ushort* dst, int len, int cn ) {
{
auto a = __riscv_vle16_v_u16m1(src0 + i, vl);
auto b = __riscv_vle16_v_u16m1(src1 + i, vl);
__riscv_vsse16_v_u16m1(dst + i*cn, sizeof(ushort)*2, a, vl);
__riscv_vsse16_v_u16m1(dst + i*cn + 1, sizeof(ushort)*2, b, vl);
__riscv_vsse16_v_u16m1(dst + i*cn, sizeof(ushort)*cn, a, vl);
__riscv_vsse16_v_u16m1(dst + i*cn + 1, sizeof(ushort)*cn, b, vl);
}
#if defined(__clang__)
#pragma clang loop vectorize(disable)
@ -167,9 +184,9 @@ static int merge16u(const ushort** src, ushort* dst, int len, int cn ) {
auto a = __riscv_vle16_v_u16m1(src0 + i, vl);
auto b = __riscv_vle16_v_u16m1(src1 + i, vl);
auto c = __riscv_vle16_v_u16m1(src2 + i, vl);
__riscv_vsse16_v_u16m1(dst + i*cn, sizeof(ushort)*3, a, vl);
__riscv_vsse16_v_u16m1(dst + i*cn + 1, sizeof(ushort)*3, b, vl);
__riscv_vsse16_v_u16m1(dst + i*cn + 2, sizeof(ushort)*3, c, vl);
__riscv_vsse16_v_u16m1(dst + i*cn, sizeof(ushort)*cn, a, vl);
__riscv_vsse16_v_u16m1(dst + i*cn + 1, sizeof(ushort)*cn, b, vl);
__riscv_vsse16_v_u16m1(dst + i*cn + 2, sizeof(ushort)*cn, c, vl);
}
#if defined(__clang__)
#pragma clang loop vectorize(disable)
@ -190,10 +207,10 @@ static int merge16u(const ushort** src, ushort* dst, int len, int cn ) {
auto b = __riscv_vle16_v_u16m1(src1 + i, vl);
auto c = __riscv_vle16_v_u16m1(src2 + i, vl);
auto d = __riscv_vle16_v_u16m1(src3 + i, vl);
__riscv_vsse16_v_u16m1(dst + i*cn, sizeof(ushort)*4, a, vl);
__riscv_vsse16_v_u16m1(dst + i*cn + 1, sizeof(ushort)*4, b, vl);
__riscv_vsse16_v_u16m1(dst + i*cn + 2, sizeof(ushort)*4, c, vl);
__riscv_vsse16_v_u16m1(dst + i*cn + 3, sizeof(ushort)*4, d, vl);
__riscv_vsse16_v_u16m1(dst + i*cn, sizeof(ushort)*cn, a, vl);
__riscv_vsse16_v_u16m1(dst + i*cn + 1, sizeof(ushort)*cn, b, vl);
__riscv_vsse16_v_u16m1(dst + i*cn + 2, sizeof(ushort)*cn, c, vl);
__riscv_vsse16_v_u16m1(dst + i*cn + 3, sizeof(ushort)*cn, d, vl);
}
#if defined(__clang__)
#pragma clang loop vectorize(disable)
@ -212,10 +229,24 @@ static int merge16u(const ushort** src, ushort* dst, int len, int cn ) {
for( ; k < cn; k += 4 )
{
const uint16_t *src0 = src[k], *src1 = src[k+1], *src2 = src[k+2], *src3 = src[k+3];
for( i = 0, j = k; i < len; i++, j += cn )
i = 0;
for( ; i <= len - vl; i += vl)
{
dst[j] = src0[i]; dst[j+1] = src1[i];
dst[j+2] = src2[i]; dst[j+3] = src3[i];
auto a = __riscv_vle16_v_u16m1(src0 + i, vl);
auto b = __riscv_vle16_v_u16m1(src1 + i, vl);
auto c = __riscv_vle16_v_u16m1(src2 + i, vl);
auto d = __riscv_vle16_v_u16m1(src3 + i, vl);
__riscv_vsse16_v_u16m1(dst + k+i*cn, sizeof(ushort)*cn, a, vl);
__riscv_vsse16_v_u16m1(dst + k+i*cn + 1, sizeof(ushort)*cn, b, vl);
__riscv_vsse16_v_u16m1(dst + k+i*cn + 2, sizeof(ushort)*cn, c, vl);
__riscv_vsse16_v_u16m1(dst + k+i*cn + 3, sizeof(ushort)*cn, d, vl);
}
for( ; i < len; i++ )
{
dst[k+i*cn] = src0[i];
dst[k+i*cn+1] = src1[i];
dst[k+i*cn+2] = src2[i];
dst[k+i*cn+3] = src3[i];
}
}
return CV_HAL_ERROR_OK;
@ -224,7 +255,7 @@ static int merge16u(const ushort** src, ushort* dst, int len, int cn ) {
#if defined __GNUC__
__attribute__((optimize("no-tree-vectorize")))
#endif
static int merge32s(const int** src, int* dst, int len, int cn ) {
inline int merge32s(const int** src, int* dst, int len, int cn ) {
int k = cn % 4 ? cn % 4 : 4;
int i, j;
if( k == 1 )
@ -294,7 +325,7 @@ static int merge32s(const int** src, int* dst, int len, int cn ) {
#if defined __GNUC__
__attribute__((optimize("no-tree-vectorize")))
#endif
static int merge64s(const int64** src, int64* dst, int len, int cn ) {
inline int merge64s(const int64** src, int64* dst, int len, int cn ) {
int k = cn % 4 ? cn % 4 : 4;
int i, j;
if( k == 1 )

23
modules/core/test/test_mat.cpp
View File

@ -1749,18 +1749,22 @@ TEST(Core_Mat_array, copyTo_roi_row)
EXPECT_EQ(5, (int)dst2[4]);
}
TEST(Core_Mat_array, SplitMerge)
typedef testing::TestWithParam< tuple<int, perf::MatType> > Core_Mat_arrays;
TEST_P(Core_Mat_arrays, SplitMerge)
{
std::array<cv::Mat, 3> src;
int cn = get<0>(GetParam());
int type = get<1>(GetParam());
std::vector<cv::Mat> src(cn);
for (size_t i = 0; i < src.size(); ++i)
{
src[i] = Mat(10, 10, CV_8U, Scalar((double)(16 * (i + 1))));
src[i] = Mat(10, 10, type, Scalar((double)(16 * (i + 1))));
}
Mat merged;
merge(src, merged);
std::array<cv::Mat, 3> dst;
std::vector<cv::Mat> dst(cn);
split(merged, dst);
for (size_t i = 0; i < dst.size(); ++i)
@ -1769,6 +1773,17 @@ TEST(Core_Mat_array, SplitMerge)
}
}
INSTANTIATE_TEST_CASE_P(/*nothing*/, Core_Mat_arrays, testing::Combine(
testing::Range(1, 9),
testing::Values(
perf::MatType(CV_8U),
perf::MatType(CV_16U),
perf::MatType(CV_32S),
perf::MatType(CV_64F)
)
)
);
TEST(Mat, regression_8680)
{
Mat_<Point2i> mat(3,1);