Merge pull request #27216 from CodeLinaro:xuezha_3rdPost

asmorkalov · web-flow · commit 3962803e7a3c · 2025-04-16T17:48:36.000+03:00
Add SVD into FastCV HAL
diff --git a/3rdparty/fastcv/include/fastcv_hal_core.hpp b/3rdparty/fastcv/include/fastcv_hal_core.hpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2024 Qualcomm Innovation Center, Inc. All rights reserved.
+ * Copyright (c) 2024-2025 Qualcomm Innovation Center, Inc. All rights reserved.
  * SPDX-License-Identifier: Apache-2.0
 */
 
@@ -32,6 +32,9 @@
 #define cv_hal_mul16s               fastcv_hal_mul16s
 #undef  cv_hal_mul32f
 #define cv_hal_mul32f               fastcv_hal_mul32f
+#undef  cv_hal_SVD32f
+#define cv_hal_SVD32f               fastcv_hal_SVD32f
+
 
 ////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
 /// @brief look-up table transform of an array.
@@ -219,4 +222,32 @@ int fastcv_hal_mul32f(
     int             height,
     double          scale);
 
+////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
+/// Performs singular value decomposition of \f$M\times N\f$(\f$M>N\f$) matrix \f$A = U*\Sigma*V^T\f$.
+///
+/// @param src      Pointer to input MxN matrix A stored in column major order.
+///                 After finish of work src will be filled with rows of U or not modified (depends of flag CV_HAL_SVD_MODIFY_A).
+/// @param src_step Number of bytes between two consequent columns of matrix A.
+/// @param w        Pointer to array for singular values of matrix A (i. e. first N diagonal elements of matrix \f$\Sigma\f$).
+/// @param u        Pointer to output MxN or MxM matrix U (size depends of flags).
+///                 Pointer must be valid if flag CV_HAL_SVD_MODIFY_A not used.
+/// @param u_step   Number of bytes between two consequent rows of matrix U.
+/// @param vt       Pointer to array for NxN matrix V^T.
+/// @param vt_step  Number of bytes between two consequent rows of matrix V^T.
+/// @param m        Number fo rows in matrix A.
+/// @param n        Number of columns in matrix A.
+/// @param flags    Algorithm options (combination of CV_HAL_SVD_FULL_UV, ...).
+////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
+int fastcv_hal_SVD32f(
+    float* src,
+    size_t src_step,
+    float* w,
+    float* u,
+    size_t u_step,
+    float* vt,
+    size_t vt_step,
+    int    m,
+    int    n,
+    int    flags);
+
 #endif
diff --git a/3rdparty/fastcv/src/fastcv_hal_core.cpp b/3rdparty/fastcv/src/fastcv_hal_core.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2024 Qualcomm Innovation Center, Inc. All rights reserved.
+ * Copyright (c) 2024-2025 Qualcomm Innovation Center, Inc. All rights reserved.
  * SPDX-License-Identifier: Apache-2.0
 */
 
@@ -399,7 +399,7 @@ int fastcv_hal_mul8u(
     int8_t sF;
 
     if(FCV_CMP_EQ(scale,1.0))              { sF =  0; }
-    else if(scale > 1.0)                    
+    else if(scale > 1.0)
     {
         if(FCV_CMP_EQ(scale,2.0))          { sF = -1; }
         else if(FCV_CMP_EQ(scale,4.0))     { sF = -2; }
@@ -471,7 +471,7 @@ int fastcv_hal_mul16s(
     int8_t sF;
 
     if(FCV_CMP_EQ(scale,1.0))              { sF =  0; }
-    else if(scale > 1.0)                    
+    else if(scale > 1.0)
     {
         if(FCV_CMP_EQ(scale,2.0))          { sF = -1; }
         else if(FCV_CMP_EQ(scale,4.0))     { sF = -2; }
@@ -571,4 +571,56 @@ int fastcv_hal_mul32f(
 
     fcvStatus status = FASTCV_SUCCESS;
     CV_HAL_RETURN(status, hal_mul32f);
+}
+
+int fastcv_hal_SVD32f(
+    float*  src,
+    size_t  src_step,
+    float*  w,
+    float*  u,
+    size_t  u_step,
+    float*  vt,
+    size_t  vt_step,
+    int     m,
+    int     n,
+    int     flags)
+{
+    if (n * sizeof(float) != src_step)
+        CV_HAL_RETURN_NOT_IMPLEMENTED("step is not supported");
+
+    INITIALIZATION_CHECK;
+
+    fcvStatus status = FASTCV_SUCCESS;
+
+    cv::Mat tmpU(m, n, CV_32F);
+    cv::Mat tmpV(n, n, CV_32F);
+
+    switch (flags)
+    {
+        case CV_HAL_SVD_NO_UV:
+        {
+            status = fcvSVDf32_v2(src, m, n, w, u, vt, (float32_t *)tmpU.data, (float32_t *)tmpV.data, false);
+            break;
+        }
+        case CV_HAL_SVD_SHORT_UV:
+        {
+            if ((n * sizeof(float) == u_step) && (n * sizeof(float) == vt_step))
+                status = fcvSVDf32_v2(src, m, n, w, u, vt, (float32_t *)tmpU.data, (float32_t *)tmpV.data, false);
+            else
+                CV_HAL_RETURN_NOT_IMPLEMENTED("step is not supported");
+            break;
+        }
+        case CV_HAL_SVD_FULL_UV:
+        {
+            if ((n * sizeof(float) == u_step) && (n * sizeof(float) == vt_step))
+                status = fcvSVDf32_v2(src, m, n, w, u, vt, (float32_t *)tmpU.data, (float32_t *)tmpV.data, true);
+            else
+                CV_HAL_RETURN_NOT_IMPLEMENTED("step is not supported");
+            break;
+        }
+        default:
+            CV_HAL_RETURN_NOT_IMPLEMENTED(cv::format("Flags:%d is not supported", flags));
+    }
+
+    CV_HAL_RETURN(status, fastcv_hal_SVD32f);
 }
