refactored suggested changes

modules/core/include/opencv2/core/hal/intrin_avx.hpp

@@ -951,7 +951,7 @@ OPENCV_HAL_IMPL_AVX_CMP_OP_64BIT(v_int64x4)
 
 #define OPENCV_HAL_IMPL_AVX_CMP_OP_FLT(_Tpvec, suffix)               \
     OPENCV_HAL_IMPL_AVX_CMP_FLT(v_eq, _CMP_EQ_OQ,  _Tpvec, suffix)   \
-    OPENCV_HAL_IMPL_AVX_CMP_FLT(v_ne, _CMP_NEQ_UQ, _Tpvec, suffix)   \
+    OPENCV_HAL_IMPL_AVX_CMP_FLT(v_ne, _CMP_NEQ_OQ, _Tpvec, suffix)   \
     OPENCV_HAL_IMPL_AVX_CMP_FLT(v_lt,  _CMP_LT_OQ,  _Tpvec, suffix)  \
     OPENCV_HAL_IMPL_AVX_CMP_FLT(v_gt,  _CMP_GT_OQ,  _Tpvec, suffix)  \
     OPENCV_HAL_IMPL_AVX_CMP_FLT(v_le, _CMP_LE_OQ,  _Tpvec, suffix)   \

modules/core/include/opencv2/core/hal/intrin_avx512.hpp

@@ -888,7 +888,7 @@ OPENCV_HAL_IMPL_AVX512_CMP_OP_INT(v_int64x8,   epi64, epi64, (int64)-1)
 
 #define OPENCV_HAL_IMPL_AVX512_CMP_OP_FLT(_Tpvec, sufcmp, sufset, tval)           \
     OPENCV_HAL_IMPL_AVX512_CMP_FLT(v_eq, _CMP_EQ_OQ,  _Tpvec, sufcmp, sufset, tval)  \
-    OPENCV_HAL_IMPL_AVX512_CMP_FLT(v_ne, _CMP_NEQ_UQ, _Tpvec, sufcmp, sufset, tval)  \
+    OPENCV_HAL_IMPL_AVX512_CMP_FLT(v_ne, _CMP_NEQ_OQ, _Tpvec, sufcmp, sufset, tval)  \
     OPENCV_HAL_IMPL_AVX512_CMP_FLT(v_lt,  _CMP_LT_OQ,  _Tpvec, sufcmp, sufset, tval) \
     OPENCV_HAL_IMPL_AVX512_CMP_FLT(v_gt,  _CMP_GT_OQ,  _Tpvec, sufcmp, sufset, tval) \
     OPENCV_HAL_IMPL_AVX512_CMP_FLT(v_le, _CMP_LE_OQ,  _Tpvec, sufcmp, sufset, tval)  \

modules/core/include/opencv2/core/hal/intrin_lasx.hpp

@@ -1026,8 +1026,8 @@ OPENCV_HAL_IMPL_LASX_CMP_OP_64BIT(v_int64x4, d)
     { return _Tpvec(__lasx_##suffix##_##ssuffix(a.val, b.val)); }
 
 #define OPENCV_HAL_IMPL_LASX_CMP_OP_FLT(_Tpvec, ssuffix)              \
-    OPENCV_HAL_IMPL_LASX_CMP_FLT(v_eq, xvfcmp_ceq,  _Tpvec, ssuffix)  \
-    OPENCV_HAL_IMPL_LASX_CMP_FLT(v_ne, xvfcmp_cune, _Tpvec, ssuffix)  \
+    OPENCV_HAL_IMPL_LASX_CMP_FLT(v_eq, xvfcmp_ceq, _Tpvec, ssuffix)   \
+    OPENCV_HAL_IMPL_LASX_CMP_FLT(v_ne, xvfcmp_cne, _Tpvec, ssuffix)   \
     OPENCV_HAL_IMPL_LASX_CMP_FLT(v_lt,  xvfcmp_clt, _Tpvec, ssuffix)  \
     OPENCV_HAL_IMPL_LASX_CMP_FLT(v_le, xvfcmp_cle, _Tpvec, ssuffix)
 

modules/core/include/opencv2/core/hal/intrin_lsx.hpp

@@ -800,8 +800,8 @@ OPENCV_HAL_IMPL_LSX_CMP_OP_64BIT(v_int64x2, d)
     { return _Tpvec(__lsx_##suffix##_##ssuffix(a.val, b.val)); }           \
 
 #define OPENCV_HAL_IMPL_LSX_CMP_OP_FLT(_Tpvec, ssuffix)                    \
-    OPENCV_HAL_IMPL_LSX_CMP_FLT(v_eq, vfcmp_ceq,  _Tpvec, ssuffix)         \
-    OPENCV_HAL_IMPL_LSX_CMP_FLT(v_ne, vfcmp_cune, _Tpvec, ssuffix)         \
+    OPENCV_HAL_IMPL_LSX_CMP_FLT(v_eq, vfcmp_ceq, _Tpvec, ssuffix)          \
+    OPENCV_HAL_IMPL_LSX_CMP_FLT(v_ne, vfcmp_cne, _Tpvec, ssuffix)          \
     OPENCV_HAL_IMPL_LSX_CMP_FLT(v_lt,  vfcmp_clt, _Tpvec, ssuffix)         \
     OPENCV_HAL_IMPL_LSX_CMP_FLT(v_le, vfcmp_cle, _Tpvec, ssuffix)          \
 

modules/core/include/opencv2/core/parallel/backend/parallel_for.tbb.hpp

@@ -27,18 +27,14 @@ using namespace ::tbb;
 #if TBB_INTERFACE_VERSION >= 8000
 static tbb::task_arena& getArena()
 {
-    // Keep the TBB object alive until process termination to avoid static
-    // destruction during process teardown on Windows.
-    static tbb::task_arena* tbbArena = new tbb::task_arena(tbb::task_arena::automatic);
-    return *tbbArena;
+    static tbb::task_arena tbbArena(tbb::task_arena::automatic);
+    return tbbArena;
 }
 #else
 static tbb::task_scheduler_init& getScheduler()
 {
-    // Keep the TBB object alive until process termination to avoid static
-    // destruction during process teardown on Windows.
-    static tbb::task_scheduler_init* tbbScheduler = new tbb::task_scheduler_init(tbb::task_scheduler_init::deferred);
-    return *tbbScheduler;
+    static tbb::task_scheduler_init tbbScheduler(tbb::task_scheduler_init::deferred);
+    return tbbScheduler;
 }
 #endif
 

modules/core/include/opencv2/core/version.hpp

@@ -8,7 +8,7 @@
 #define CV_VERSION_MAJOR    5
 #define CV_VERSION_MINOR    1
 #define CV_VERSION_REVISION 0
-#define CV_VERSION_STATUS   ""
+#define CV_VERSION_STATUS   "-pre"
 
 #define CVAUX_STR_EXP(__A)  #__A
 #define CVAUX_STR(__A)      CVAUX_STR_EXP(__A)

modules/core/src/arithm.simd.hpp

@@ -25,7 +25,6 @@
 #include <algorithm>
 #include <cmath>
 #include <cstring>
-#include <type_traits>
 
 namespace cv {
 
@@ -384,7 +383,7 @@ struct EwMul {
 #endif
     template<typename V, typename S> static V vec(const V& a, const V& b, const S&) { return vec(a, b); }
     template<typename V> static V preproc(const V& a, const V& s) { return v_mul(a, s); }
-    template<typename W, typename ST> static W scl(W a, W b, ST s) { return a * s * b; }
+    template<typename W, typename ST> static W scl(W a, W b, ST s) { return a * b * s; }
 };
 // div has two variants by the COMMON INPUT type (matching cv::'s per-type kernel choice): integer
 // inputs guard divide-by-zero -> 0 (cv:: iscalar_div); float inputs do NOT guard (cv:: fscalar_div,
@@ -574,11 +573,6 @@ static int scalarBinaryKernel(const void* src0_, size_t s0y, size_t s0x,
     s1y /= sizeof(T);
     dsty /= sizeof(Tr);
     CV_Assert((s0x|s1x) == 1u || (s0x|s1x) + (size_t)width == 1u);
-    if constexpr (std::is_same_v<Op, EwMul>) {
-        if ((s1x == 0 && s0x != 0) || (s1y == 0 && s0y != 0 && s0x == s1x)) {
-            std::swap(src0_, src1_); std::swap(s0y, s1y); std::swap(s0x, s1x);
-        }
-    }
 
     const T* src0 = (const T*)src0_;
     const T* src1 = (const T*)src1_;
@@ -922,11 +916,6 @@ static int vecBinaryKernel(const void* src0_, size_t s0y, size_t s0x,
     dsty /= sizeof(Tr);
 
     CV_Assert((s0x|s1x) == 1u || (s0x|s1x) + (size_t)width == 1u);
-    if constexpr (std::is_same_v<Op, EwMul>) {
-        if ((s1x == 0 && s0x != 0) || (s1y == 0 && s0y != 0 && s0x == s1x)) {
-            std::swap(src0_, src1_); std::swap(s0y, s1y); std::swap(s0x, s1x);
-        }
-    }
 
     const T* src0 = (const T*)src0_;
     const T* src1 = (const T*)src1_;

modules/core/src/arithm_expr.cpp

@@ -1231,122 +1231,6 @@ static constexpr int LOCAL_HDRS = 3;
 static constexpr int LOCAL_CONSTS = 4;
 static constexpr int LOCAL_OPS = 16;
 
-#if defined(__GNUC__)
-#define CV_EW_NOINLINE __attribute__((noinline))
-#elif defined(_MSC_VER)
-#define CV_EW_NOINLINE __declspec(noinline)
-#else
-#define CV_EW_NOINLINE
-#endif
-
-// Outlined from TExpr::exec() so the ~4KB inline scratchBuf is not hoisted into TExpr::exec()'s
-// stack frame when execution takes the broadcastOp path instead.
-CV_EW_NOINLINE static void execFastStripWithTemps(
-    const TExpr& expr, const Mat* const* inputs, Mat* outputs, size_t total)
-{
-    const int ninsn = (int)expr.prog.size();
-    const int totalEsz = expr.bufEszPrefix[expr.nbuffers];
-    const int wf0 = std::min((int)total, expr.capElems);
-    const size_t region = alignSize((size_t)wf0, 8);
-    // Scratch for the temp buffers. AutoBuffer no longer value-inits its tail, so a fresh per-call
-    // buffer is free (we only WRITE to it); the inline 4KB covers the L1-capped size, heap backs
-    // the rare larger case. uint64_t is used for memory alignment on the worst case.
-    AutoBuffer<uint64_t, (4*1024 + 256)/sizeof(uint64_t)> scratchBuf(
-        divUp((size_t)totalEsz * region, (unsigned)sizeof(uint64_t)));
-    uchar* scratch = (uchar*)scratchBuf.data();
-    for (int x0 = 0; x0 < (int)total; x0 += wf0)
-    {
-        const int wf = std::min(wf0, (int)total - x0);
-        auto ptrOf = [&](int s) -> void* {
-            if (s <= 0) return nullptr;
-            const TExpr::Arg& ai = expr.arginfo[s];
-            size_t esz = CV_ELEM_SIZE1(ai.depth);
-            if (ai.kind == TExpr::INPUT)  return (uchar*)inputs[ai.index]->data  + (size_t)x0 * esz;
-            if (ai.kind == TExpr::OUTPUT) return (uchar*)outputs[ai.index].data + (size_t)x0 * esz;
-            if (ai.kind != TExpr::TEMP)   return nullptr;                                         // NONE: moved-from
-            const int b = expr.bufferOfTemp.empty() ? ai.index : expr.bufferOfTemp[ai.index];     // TEMP
-            return scratch + (size_t)expr.bufEszPrefix[b] * region;                               // fragment-local
-        };
-        for (int n = 0; n < ninsn; n++)
-        {
-            const TExpr::Insn& ins = expr.prog[n];
-            runInsn(ins, ptrOf(ins.arg0), 0, 1, ptrOf(ins.arg1), 0, 1,
-                    ptrOf(ins.arg2), 0, 1, ptrOf(ins.result), 0, wf, 1);
-        }
-    }
-}
-
-// Outlined from the broadcastOp body so single-op calls (nbuffers == 0) do not reserve ~4KB of
-// unused tstoreBuf stack space in the closure frame.
-CV_EW_NOINLINE static void execTileWithTemps(
-    const EwBody& bc, const TExpr::Arg* arginfo, const BroadcastOp::Tile& tile)
-{
-    using BrSlice = BroadcastOp::Slice;
-    const TExpr::Insn* insns = bc.prog;
-    const int* const smap  = bc.slotMap;
-    const signed char* const skind = bc.slotKind;
-    const int nin = bc.ninsn, nsl = bc.nslots;
-    const int w = tile.width, h = tile.height;
-
-    // Run the program over L1-sized 2D blocks so the intermediates stay in cache. ONE rule covers
-    // every tile shape: keep the strip as WIDE as fits L1 (long inner loop => good SIMD + hits the
-    // kernels' width-specific branches), then add as many rows as still fit (bw*bh <= capElems).
-    // For the dominant 1D tile (w huge, h==1) this is a column strip; for a tall-thin tile
-    // (w==channels, h huge - e.g. a masked op) it instead splits the LONG axis (height), so the
-    // temp block stays contiguous and the kernels keep full width. Each temp buffer occupies
-    // [bufEszPrefix[buf]*region, ...) bytes in tstore; region = alignSize(bw*bh, 8).
-    const int* const eszPrefix = bc.bufEszPrefix;
-    const int bw = std::min(w, bc.capElems);
-    const int bh = std::min(h, std::max(1, bc.capElems / std::max(1, bw)));
-    const size_t region = alignSize((size_t)bw * bh, 8);
-    // uint64_t elements (see the note on scratchBuf above): keeps the block-local temp store
-    // 8-byte aligned, which the 64-bit-lane kernels require on 32-bit ABIs.
-    AutoBuffer<uint64_t, (4*1024 + 256)/sizeof(uint64_t)> tstoreBuf(
-        divUp((size_t)eszPrefix[bc.nbuffers] * region, (unsigned)sizeof(uint64_t)));  // inline (<= ~4KB)
-    uchar* tstore = (uchar*)tstoreBuf.data();
-
-    AutoBuffer<BrSlice, LOCAL_OPS> args(nsl);
-    for (int y0 = 0; y0 < h; y0 += bh)
-    {
-        const int hf = std::min(bh, h - y0);
-        for (int x0 = 0; x0 < w; x0 += bw)
-        {
-            const int wf = std::min(bw, w - x0);
-            for (int s = 1; s < nsl; s++)
-            {
-                BrSlice& a = args[s];
-                const int k = skind[s];
-                if (k == TExpr::NONE) continue;
-                size_t esz = CV_ELEM_SIZE1(arginfo[s].depth);
-                if (k == TExpr::TEMP)                      // contiguous block-local buffer
-                {
-                    a.ptr = tstore + (size_t)eszPrefix[smap[s]] * region;
-                    a.stepy = (size_t)wf*esz; a.stepx = 1;
-                }
-                else                                       // array operand (incl. 0-dim consts): this
-                                                           // block of the broadcast slice
-                {
-                    const BrSlice& sl = tile.slices[smap[s]];
-                    a.ptr = (uchar*)sl.ptr +
-                        ((size_t)y0 * sl.stepy + (size_t)x0 * sl.stepx) * esz;
-                    a.stepy = sl.stepy*esz; a.stepx = sl.stepx;
-                }
-            }
-
-            for (int n = 0; n < nin; n++)
-            {
-                const TExpr::Insn& ins = insns[n];
-                const BrSlice& a0 = args[ins.arg0]; const BrSlice& a1 = args[ins.arg1];
-                const BrSlice& a2 = args[ins.arg2]; const BrSlice& rr = args[ins.result];
-                runInsn(ins, a0.ptr, a0.stepy, a0.stepx, a1.ptr, a1.stepy, a1.stepx,
-                        a2.ptr, a2.stepy, a2.stepx, (void*)rr.ptr, rr.stepy, wf, hf);
-            }
-        }
-    }
-}
-
-#undef CV_EW_NOINLINE
-
 void TExpr::exec(const Mat* const* inputs, Mat* outputs)
 {
     using BrTile = BroadcastOp::Tile;
@@ -1417,7 +1301,38 @@ void TExpr::exec(const Mat* const* inputs, Mat* outputs)
             }
 
             // Temps present: walk the strip in L1-sized fragments so the intermediates stay hot.
-            execFastStripWithTemps(*this, inputs, outputs, total);
+            // Byte layout (bufEszPrefix) and the fragment cap (capElems) both come from compile();
+            // scratch is totalEsz*wf0 (<= ~4KB).
+            const int totalEsz = bufEszPrefix[nbuffers];
+            const int wf0 = std::min((int)total, capElems);
+
+            const size_t region = alignSize((size_t)wf0, 8);
+            // Scratch for the temp buffers. AutoBuffer no longer value-inits its tail, so a fresh per-call
+            // buffer is free (we only WRITE to it); the inline 4KB covers the L1-capped size, heap backs
+            // the rare larger case. uint64_t is used for memory alignment on the worst case.
+            AutoBuffer<uint64_t, (4*1024 + 256)/sizeof(uint64_t)> scratchBuf(
+                divUp((size_t)totalEsz * region, (unsigned)sizeof(uint64_t)));
+            uchar* scratch = (uchar*)scratchBuf.data();
+            for (int x0 = 0; x0 < (int)total; x0 += wf0)
+            {
+                const int wf = std::min(wf0, (int)total - x0);
+                auto ptrOf = [&](int s) -> void* {
+                    if (s <= 0) return nullptr;
+                    const Arg& ai = arginfo[s];
+                    size_t esz = CV_ELEM_SIZE1(ai.depth);
+                    if (ai.kind == INPUT)  return (uchar*)inputs[ai.index]->data  + (size_t)x0 * esz;
+                    if (ai.kind == OUTPUT) return (uchar*)outputs[ai.index].data + (size_t)x0 * esz;
+                    if (ai.kind != TEMP)   return nullptr;                                    // NONE: moved-from
+                    const int b = bufferOfTemp.empty() ? ai.index : bufferOfTemp[ai.index];  // TEMP
+                    return scratch + (size_t)bufEszPrefix[b] * region;                        // fragment-local
+                };
+                for (int n = 0; n < ninsn; n++)
+                {
+                    const Insn& ins = prog[n];
+                    runInsn(ins, ptrOf(ins.arg0), 0, 1, ptrOf(ins.arg1), 0, 1,
+                            ptrOf(ins.arg2), 0, 1, ptrOf(ins.result), 0, wf, 1);
+                }
+            }
             return;
         }
     }
@@ -1551,38 +1466,70 @@ void TExpr::exec(const Mat* const* inputs, Mat* outputs)
     body.nbuffers = nbuffers;
     body.capElems = capElems;
 
-    broadcastOp(arr.data(), narr, [&bc=body, this](const BrTile& tile)
+    broadcastOp(arr.data(), narr, [&](const BrTile& tile)
     {
-        if (bc.nbuffers != 0)
-        {
-            execTileWithTemps(bc, arginfo.data(), tile);
-            return;
-        }
-
-        // No temps: one block over the whole tile, no scratch buffer on the stack, and each
-        // operand slot is re-pointed once straight at its tile slice.
+        EwBody& bc = body;
         const TExpr::Insn* insns = bc.prog;            // hot fields -> locals (registers/stack)
         const int* const smap  = bc.slotMap;
         const signed char* const skind = bc.slotKind;
         const int nin = bc.ninsn, nsl = bc.nslots;
         const int w = tile.width, h = tile.height;
 
-        AutoBuffer<BrSlice, LOCAL_OPS> args(nsl);
-        for (int s = 1; s < nsl; s++)
-        {
-            if (skind[s] == TExpr::NONE) continue;
-            const BrSlice& sl = tile.slices[smap[s]];
-            size_t esz = CV_ELEM_SIZE1(arginfo[s].depth);
-            args[s] = { sl.ptr, sl.stepy * esz, sl.stepx };
-        }
+        // Run the program over L1-sized 2D blocks so the intermediates stay in cache. ONE rule covers
+        // every tile shape: keep the strip as WIDE as fits L1 (long inner loop => good SIMD + hits the
+        // kernels' width-specific branches), then add as many rows as still fit (bw*bh <= capElems).
+        // For the dominant 1D tile (w huge, h==1) this is a column strip; for a tall-thin tile
+        // (w==channels, h huge - e.g. a masked op) it instead splits the LONG axis (height), so the
+        // temp block stays contiguous and the kernels keep full width. With NO temps capElems==INT_MAX
+        // => bw=w, bh=h: one block over the whole tile, region temp store is empty (zero bytes), and
+        // each operand slot is re-pointed once straight at its tile slice. Each temp buffer occupies
+        // [bufEszPrefix[buf]*region, ...) bytes in tstore; region = bw*bh.
+        const int* const eszPrefix = bc.bufEszPrefix;
+        const int bw = std::min(w, bc.capElems);
+        const int bh = std::min(h, std::max(1, bc.capElems / std::max(1, bw)));
+        const size_t region = alignSize((size_t)bw * bh, 8);
+        // uint64_t elements (see the note on scratchBuf above): keeps the block-local temp store
+        // 8-byte aligned, which the 64-bit-lane kernels require on 32-bit ABIs.
+        AutoBuffer<uint64_t, (4*1024 + 256)/sizeof(uint64_t)> tstoreBuf(
+            divUp((size_t)eszPrefix[bc.nbuffers] * region, (unsigned)sizeof(uint64_t)));  // inline (<= ~4KB)
+        uchar* tstore = (uchar*)tstoreBuf.data();
 
-        for (int n = 0; n < nin; n++)
+        AutoBuffer<BrSlice, LOCAL_OPS> args(nsl);
+        for (int y0 = 0; y0 < h; y0 += bh)
         {
-            const TExpr::Insn& ins = insns[n];
-            const BrSlice& a0 = args[ins.arg0]; const BrSlice& a1 = args[ins.arg1];
-            const BrSlice& a2 = args[ins.arg2]; const BrSlice& rr = args[ins.result];
-            runInsn(ins, a0.ptr, a0.stepy, a0.stepx, a1.ptr, a1.stepy, a1.stepx,
-                    a2.ptr, a2.stepy, a2.stepx, (void*)rr.ptr, rr.stepy, w, h);
+            const int hf = std::min(bh, h - y0);
+            for (int x0 = 0; x0 < w; x0 += bw)
+            {
+                const int wf = std::min(bw, w - x0);
+                for (int s = 1; s < nsl; s++)
+                {
+                    BrSlice& a = args[s];
+                    const int k = skind[s];
+                    size_t esz = CV_ELEM_SIZE1(arginfo[s].depth);
+                    if (k == TExpr::TEMP)                      // contiguous block-local buffer
+                    {
+                        a.ptr = tstore + (size_t)eszPrefix[smap[s]] * region;
+                        a.stepy = (size_t)wf*esz; a.stepx = 1;
+                    }
+                    else                                    // array operand (incl. 0-dim consts): this
+                                                            // block of the broadcast slice
+                    {
+                        const BrSlice& sl = tile.slices[smap[s]];
+                        a.ptr = (uchar*)sl.ptr +
+                            ((size_t)y0 * sl.stepy + (size_t)x0 * sl.stepx) * esz;
+                        a.stepy = sl.stepy*esz; a.stepx = sl.stepx;
+                    }
+                }
+
+                for (int n = 0; n < nin; n++)
+                {
+                    const TExpr::Insn& ins = insns[n];
+                    const BrSlice& a0 = args[ins.arg0]; const BrSlice& a1 = args[ins.arg1];
+                    const BrSlice& a2 = args[ins.arg2]; const BrSlice& rr = args[ins.result];
+                    runInsn(ins, a0.ptr, a0.stepy, a0.stepx, a1.ptr, a1.stepy, a1.stepx,
+                            a2.ptr, a2.stepy, a2.stepx, (void*)rr.ptr, rr.stepy, wf, hf);
+                }
+            }
         }
     }, true, nstripes);
 }

modules/core/src/parallel.cpp

@@ -465,17 +465,9 @@ namespace {
 
 #if defined HAVE_TBB
     #if TBB_INTERFACE_VERSION >= 8000
-        static tbb::task_arena& getTbbArena() {
-            // Keep the TBB object alive until process termination to avoid static
-            // destruction during process teardown on Windows.
-            CV_SINGLETON_LAZY_INIT_REF(tbb::task_arena, new tbb::task_arena(tbb::task_arena::automatic));
-        }
+        static tbb::task_arena tbbArena(tbb::task_arena::automatic);
     #else
-        static tbb::task_scheduler_init& getTbbScheduler() {
-            // Keep the TBB object alive until process termination to avoid static
-            // destruction during process teardown on Windows.
-            CV_SINGLETON_LAZY_INIT_REF(tbb::task_scheduler_init, new tbb::task_scheduler_init(tbb::task_scheduler_init::deferred));
-        }
+        static tbb::task_scheduler_init tbbScheduler(tbb::task_scheduler_init::deferred);
     #endif
 #elif defined HAVE_HPX
 // nothing for HPX
@@ -590,7 +582,7 @@ static void parallel_for_impl(const cv::Range& range, const cv::ParallelLoopBody
 #if defined HAVE_TBB
 
 #if TBB_INTERFACE_VERSION >= 8000
-        getTbbArena().execute(pbody);
+        tbbArena.execute(pbody);
 #else
         pbody();
 #endif
@@ -659,13 +651,13 @@ int getNumThreads(void)
 #if defined HAVE_TBB
 
 #if TBB_INTERFACE_VERSION >= 9100
-    return getTbbArena().max_concurrency();
+    return tbbArena.max_concurrency();
 #elif TBB_INTERFACE_VERSION >= 8000
     return numThreads > 0
         ? numThreads
         : tbb::task_scheduler_init::default_num_threads();
 #else
-    return getTbbScheduler().is_active()
+    return tbbScheduler.is_active()
            ? numThreads
            : tbb::task_scheduler_init::default_num_threads();
 #endif
@@ -743,11 +735,11 @@ void setNumThreads( int threads_ )
 #ifdef HAVE_TBB
 
 #if TBB_INTERFACE_VERSION >= 8000
-    if(getTbbArena().is_active()) getTbbArena().terminate();
-    if(threads > 0) getTbbArena().initialize(threads);
+    if(tbbArena.is_active()) tbbArena.terminate();
+    if(threads > 0) tbbArena.initialize(threads);
 #else
-    if(getTbbScheduler().is_active()) getTbbScheduler().terminate();
-    if(threads > 0) getTbbScheduler().initialize(threads);
+    if(tbbScheduler.is_active()) tbbScheduler.terminate();
+    if(threads > 0) tbbScheduler.initialize(threads);
 #endif
 
 #elif defined HAVE_HPX

modules/core/src/persistence.cpp

@@ -1837,10 +1837,6 @@ ushort FileStorage::Impl::Base64Decoder::getUInt16() {
     return val;
 }
 
-unsigned FileStorage::Impl::Base64Decoder::getUInt32() {
-    return (unsigned)getInt32();
-}
-
 int FileStorage::Impl::Base64Decoder::getInt32() {
     size_t sz = decoded.size();
     if (ofs + 4 > sz && !readMore(4))
@@ -1918,7 +1914,7 @@ char *FileStorage::Impl::parseBase64(char *ptr, int indent, FileNode &collection
                         ival = (short) base64decoder.getUInt16();
                         break;
                     case CV_32U:
-                        ival = base64decoder.getUInt32();
+                        ival = base64decoder.getInt32();
                         break;
                     case CV_32S:
                         ival = base64decoder.getInt32();
@@ -2747,7 +2743,7 @@ FileNodeIterator& FileNodeIterator::readRaw( const String& fmt, void* _data0, si
                     FileNode node = *(*this);
                     if( node.isInt() )
                     {
-                        int64_t ival = (int64_t)node;
+                        int64_t ival = static_cast<int64_t>(elem_size == 8 ? (int64_t)node : (int)node);
                         switch( elem_type )
                         {
                         case CV_8U:

modules/core/src/persistence_impl.hpp

@@ -151,8 +151,6 @@ public:
 
         ushort getUInt16();
 
-        unsigned getUInt32();
-
         int getInt32();
 
         int64_t getInt64();

modules/core/test/test_arithm.cpp

@@ -5129,65 +5129,4 @@ TEST_P(Core_PatchNaNs, replaces_only_nans)
 
 INSTANTIATE_TEST_CASE_P(/**/, Core_PatchNaNs, testing::Values(CV_16F, CV_16BF, CV_32F, CV_64F));
 
-TEST(Core_Multiply, scale_simd_scalar_consistency)
-{
-    const float big = std::ldexp(1.0f, 80);    // exact power of 2 (< FLT_MAX = 2^128)
-    const float small = std::ldexp(1.0f, -80);
-
-    // Two contiguous inputs: (a * s) * b = 2^80 vs (a * b) * s = +inf.
-    {
-        Mat a(1, 65, CV_32F, Scalar(big));
-        Mat b(1, 65, CV_32F, Scalar(big));
-        cv::multiply(a, b, a, small);   // in-place -> tail element 64 uses EwMul::scl
-        EXPECT_EQ(a.at<float>(0),  big);
-        EXPECT_EQ(a.at<float>(64), big);
-    }
-
-    // src1 scalar/column broadcast (s1x == 0): SIMD folds scale into src1, computing (b * s) * a = 2^80;
-    // without canonicalizing src1 onto src0, the scalar tail would compute (a * s) * b = +inf.
-    {
-        Mat a(1, 65, CV_32F, Scalar(big));
-        Mat b(1, 1, CV_32F, Scalar(small));
-        cv::multiply(a, b, a, big);
-        EXPECT_EQ(a.at<float>(0),  big);
-        EXPECT_EQ(a.at<float>(64), big);
-    }
-    {
-        Mat a(1, 65, CV_32F, Scalar(big));
-        cv::multiply(a, Scalar(small), a, big);
-        EXPECT_EQ(a.at<float>(0),  big);
-        EXPECT_EQ(a.at<float>(64), big);
-    }
-
-    // src1 short-row broadcast (s1y == 0, width <= 4): short-row SIMD processes full row blocks and
-    // leaves row 64 for the scalar loop.
-    {
-        Mat a(65, 3, CV_32F, Scalar(big));
-        Mat b(1, 3, CV_32F, Scalar(small));
-        cv::multiply(a, b, a, big);
-        EXPECT_EQ(a.at<float>(0, 0),  big);
-        EXPECT_EQ(a.at<float>(64, 0), big);
-    }
-
-    // src1 wide-row broadcast (s1y == 0, width > 4):
-    {
-        Mat a(2, 65, CV_32F, Scalar(big));
-        Mat b(1, 65, CV_32F, Scalar(small));
-        cv::multiply(a, b, a, big);
-        EXPECT_EQ(a.at<float>(0, 0),  big);
-        EXPECT_EQ(a.at<float>(1, 64), big);
-    }
-
-    // CV_64F broadcast (covers v_float64 SIMD and scalarBinaryKernel fallback):
-    {
-        const double big64 = std::ldexp(1.0, 600);
-        const double small64 = std::ldexp(1.0, -600);
-        Mat a(1, 65, CV_64F, Scalar(big64));
-        Mat b(1, 1, CV_64F, Scalar(small64));
-        cv::multiply(a, b, a, big64);
-        EXPECT_EQ(a.at<double>(0),  big64);
-        EXPECT_EQ(a.at<double>(64), big64);
-    }
-}
-
 }} // namespace

modules/core/test/test_intrin_utils.hpp

@@ -726,38 +726,6 @@ template<typename R> struct TheTest
             EXPECT_EQ(dataA[i] <  dataB[i], resG[i] != 0);
             EXPECT_EQ(dataA[i] <= dataB[i], resH[i] != 0);
         }
-
-        if (std::numeric_limits<LaneType>::has_quiet_NaN || std::is_same<LaneType, hfloat>::value)
-        {
-            const LaneType qnan = (LaneType)std::numeric_limits<float>::quiet_NaN();
-            dataA[0] = qnan;
-            if (VTraits<R>::vlanes() > 1)
-            {
-                dataA[1] = qnan;
-                dataB[1] = qnan;
-            }
-            if (VTraits<R>::vlanes() > 2)
-                dataB[2] = qnan;
-            a = dataA; b = dataB;
-
-            resC = (v_eq(a, b));
-            resD = (v_ne(a, b));
-            resE = (v_gt(a, b));
-            resF = (v_ge(a, b));
-            resG = (v_lt(a, b));
-            resH = (v_le(a, b));
-
-            for (int i = 0; i < VTraits<R>::vlanes(); ++i)
-            {
-                SCOPED_TRACE(cv::format("nan i=%d", i));
-                EXPECT_EQ(dataA[i] == dataB[i], resC.as_int(i) != 0);
-                EXPECT_EQ(dataA[i] != dataB[i], resD.as_int(i) != 0);
-                EXPECT_EQ(dataA[i] >  dataB[i], resE.as_int(i) != 0);
-                EXPECT_EQ(dataA[i] >= dataB[i], resF.as_int(i) != 0);
-                EXPECT_EQ(dataA[i] <  dataB[i], resG.as_int(i) != 0);
-                EXPECT_EQ(dataA[i] <= dataB[i], resH.as_int(i) != 0);
-            }
-        }
         return *this;
     }
 

modules/core/test/test_io.cpp

@@ -2662,28 +2662,4 @@ TEST(Core_InputOutput, filestorage_read_uppercase_exponent)
     }
 }
 
-// CV_32U values above INT_MAX must survive saving and loading, in every format
-TEST(Core_InputOutput, filestorage_32u_large_values)
-{
-    const unsigned values[] = { 0u, 1u, 2147483647u, 2147483648u, 4294967295u };
-    const Mat src(1, 5, CV_32U, (void*)values);
-
-    for (const char* ext : { ".yml", ".xml", ".json" })
-    {
-        for (int base64 : { 0, (int)FileStorage::WRITE_BASE64 })
-        {
-            FileStorage writer(ext, FileStorage::WRITE | FileStorage::MEMORY | base64);
-            writer << "m" << src;
-            const std::string text = writer.releaseAndGetString();
-
-            FileStorage reader(text, FileStorage::READ | FileStorage::MEMORY);
-            Mat dst;
-            reader["m"] >> dst;
-
-            EXPECT_EQ(CV_32U, dst.type()) << ext << " base64=" << base64;
-            EXPECT_EQ(0, cv::norm(src, dst, NORM_INF)) << ext << " base64=" << base64;
-        }
-    }
-}
-
 }} // namespace

modules/dnn/perf/perf_net.cpp

@@ -28,6 +28,10 @@ public:
     {
         backend = (dnn::Backend)(int)get<0>(GetParam());
         target = (dnn::Target)(int)get<1>(GetParam());
+        // OPENCV_DNN_PERF_PIN_INPUTS=1 pins every test's inputs, to compare with the default unpinned run
+        static const bool pinAll = utils::getConfigurationParameterBool("OPENCV_DNN_PERF_PIN_INPUTS", false);
+        if (pinAll)
+            pinnedInput = "*";
     }
 
     void processNet(std::string weights, std::string proto,
@@ -36,10 +40,14 @@ public:
         if (!proto.empty())
             proto = findDataFile(proto);
         net = readNet(weights, proto);
-        // "*" pins every named input to the shape it is fed
-        if (pinnedInput == "*")
-            for (auto& inp : inputs)
-                net.setInputShape(std::get<1>(inp), shape(std::get<0>(inp)));
+        // "*" pins every input to the shape it is fed; an unnamed one is the graph input at its position
+        if (pinnedInput == "*" && net.getMainGraph()) {
+            const std::vector<Arg>& graphInputs = net.getMainGraph()->inputs();
+            for (size_t i = 0; i < inputs.size() && i < graphInputs.size(); i++) {
+                const std::string& name = std::get<1>(inputs[i]);
+                net.setInputShape(name.empty() ? net.argName(graphInputs[i]) : name, shape(std::get<0>(inputs[i])));
+            }
+        }
         else if (!pinnedInput.empty())
             net.setInputShape(pinnedInput, shape(std::get<0>(inputs[0])));
         // Set multiple inputs

modules/dnn/src/cuda/activations.cu

@@ -347,6 +347,11 @@ template void reciprocal<__half>(const Stream&, Span<__half>, View<__half>);
 
 template void relu<float>(const Stream&, Span<float>, View<float>, float);
 template void clipped_relu<float>(const Stream&, Span<float>, View<float>, float, float);
+// integer Clip (e.g. index arithmetic clamped to a valid range)
+template void clipped_relu<int8_t>(const Stream&, Span<int8_t>, View<int8_t>, int8_t, int8_t);
+template void clipped_relu<uint8_t>(const Stream&, Span<uint8_t>, View<uint8_t>, uint8_t, uint8_t);
+template void clipped_relu<int32_t>(const Stream&, Span<int32_t>, View<int32_t>, int32_t, int32_t);
+template void clipped_relu<int64_t>(const Stream&, Span<int64_t>, View<int64_t>, int64_t, int64_t);
 template void tanh<float>(const Stream&, Span<float>, View<float>);
 template void swish<float>(const Stream&, Span<float>, View<float>);
 template void mish<float>(const Stream&, Span<float>, View<float>);

modules/dnn/src/cuda/argmax.cu

@@ -0,0 +1,100 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+// Copyright (C) 2026, BigVision LLC, all rights reserved.
+// Third party copyrights are property of their respective owners.
+#include <cuda_runtime.h>
+#include <cuda_fp16.h>
+
+#include "types.hpp"
+#include "grid_stride_range.hpp"
+#include "execution.hpp"
+
+#include "../cuda4dnn/csl/stream.hpp"
+#include "../cuda4dnn/csl/span.hpp"
+
+#include "../cuda4dnn/kernels/argmax.hpp"
+
+#include <cstdint>
+
+using namespace cv::dnn::cuda4dnn::csl;
+using namespace cv::dnn::cuda4dnn::csl::device;
+
+namespace cv { namespace dnn { namespace cuda4dnn { namespace kernels {
+
+    namespace raw {
+        template <class T, bool IsArgMax, bool SelectLast>
+        __global__ void arg_min_max(View<T> input, Span<std::int64_t> output, size_type num_outputs, int axis_size, int inner_size) {
+            for (auto id : grid_stride_range(num_outputs)) {
+                int outer_index = id / inner_size;
+                int inner_index = id % inner_size;
+                std::int64_t base = (static_cast<std::int64_t>(outer_index) * axis_size) * inner_size + inner_index;
+
+                const T* p_input = input.data().get();
+                T best_value = p_input[base];
+                std::int64_t best_index = 0;
+                for (int k = 1; k < axis_size; ++k) {
+                    const T value = p_input[base + static_cast<std::int64_t>(k) * inner_size];
+                    if (IsArgMax) {
+                        if (SelectLast ? value >= best_value : value > best_value) {
+                            best_value = value;
+                            best_index = k;
+                        }
+                    } else {
+                        if (SelectLast ? value <= best_value : value < best_value) {
+                            best_value = value;
+                            best_index = k;
+                        }
+                    }
+                }
+
+                output[id] = best_index;
+            }
+        }
+    }
+
+    template <class T>
+    void arg_min_max(const Stream& stream,
+        Span<std::int64_t> output, View<T> input,
+        int outer_size, int axis_size, int inner_size, bool is_argmax, bool select_last_index)
+    {
+        if (axis_size <= 0 || outer_size <= 0 || inner_size <= 0)
+            return;
+
+        const size_type num_outputs = static_cast<size_type>(outer_size) * inner_size;
+        if (num_outputs == 0)
+            return;
+
+        if (is_argmax) {
+            if (select_last_index) {
+                auto kernel = raw::arg_min_max<T, true, true>;
+                auto policy = make_policy(kernel, num_outputs, 0, stream);
+                launch_kernel(kernel, policy, input, output, num_outputs, axis_size, inner_size);
+            } else {
+                auto kernel = raw::arg_min_max<T, true, false>;
+                auto policy = make_policy(kernel, num_outputs, 0, stream);
+                launch_kernel(kernel, policy, input, output, num_outputs, axis_size, inner_size);
+            }
+        } else {
+            if (select_last_index) {
+                auto kernel = raw::arg_min_max<T, false, true>;
+                auto policy = make_policy(kernel, num_outputs, 0, stream);
+                launch_kernel(kernel, policy, input, output, num_outputs, axis_size, inner_size);
+            } else {
+                auto kernel = raw::arg_min_max<T, false, false>;
+                auto policy = make_policy(kernel, num_outputs, 0, stream);
+                launch_kernel(kernel, policy, input, output, num_outputs, axis_size, inner_size);
+            }
+        }
+    }
+
+#if !defined(__CUDA_ARCH__) || (__CUDA_ARCH__ >= 530)
+    template void arg_min_max(const Stream&, Span<std::int64_t>, View<__half>, int, int, int, bool, bool);
+#endif
+    template void arg_min_max(const Stream&, Span<std::int64_t>, View<float>, int, int, int, bool, bool);
+    template void arg_min_max(const Stream&, Span<std::int64_t>, View<int8_t>, int, int, int, bool, bool);
+    template void arg_min_max(const Stream&, Span<std::int64_t>, View<uint8_t>, int, int, int, bool, bool);
+    template void arg_min_max(const Stream&, Span<std::int64_t>, View<int32_t>, int, int, int, bool, bool);
+    template void arg_min_max(const Stream&, Span<std::int64_t>, View<int64_t>, int, int, int, bool, bool);
+
+}}}} /* namespace cv::dnn::cuda4dnn::kernels */

modules/dnn/src/cuda/cast.cu

@@ -53,6 +53,44 @@ namespace cv { namespace dnn { namespace cuda4dnn { namespace kernels {
             __device__ static std::int64_t apply(float v) { return __float2ll_rz(v); }
         };
 
+        // 8-bit outputs saturate like the CPU's saturate_cast; integer sources need no rounding
+        template <class TIn>
+        struct Converter<std::uint8_t, TIn> {
+            __device__ static std::uint8_t apply(TIn v) {
+                const std::int64_t x = static_cast<std::int64_t>(v);
+                return static_cast<std::uint8_t>(x < 0 ? 0 : (x > 255 ? 255 : x));
+            }
+        };
+
+        template <class TIn>
+        struct Converter<std::int8_t, TIn> {
+            __device__ static std::int8_t apply(TIn v) {
+                const std::int64_t x = static_cast<std::int64_t>(v);
+                return static_cast<std::int8_t>(x < -128 ? -128 : (x > 127 ? 127 : x));
+            }
+        };
+
+        // float truncates toward zero first, as the CPU Cast does; NaN saturates like saturate_cast(INT_MIN)
+        template <>
+        struct Converter<std::uint8_t, float> {
+            __device__ static std::uint8_t apply(float v) {
+                if (isnan(v))
+                    return 0;
+                const float t = truncf(v);
+                return static_cast<std::uint8_t>(t < 0.f ? 0.f : (t > 255.f ? 255.f : t));
+            }
+        };
+
+        template <>
+        struct Converter<std::int8_t, float> {
+            __device__ static std::int8_t apply(float v) {
+                if (isnan(v))
+                    return -128;
+                const float t = truncf(v);
+                return static_cast<std::int8_t>(t < -128.f ? -128.f : (t > 127.f ? 127.f : t));
+            }
+        };
+
         template <class TOut, class TIn>
         __global__ void cast(Span<TOut> output, View<TIn> input) {
             for (auto i : grid_stride_range(output.size()))
@@ -87,4 +125,19 @@ namespace cv { namespace dnn { namespace cuda4dnn { namespace kernels {
     template void cast(const Stream&, Span<float>, View<std::int64_t>);
     template void cast(const Stream&, Span<float>, View<float>);
 
+#define CV_CUDA_CAST_8BIT(T8) \
+    template void cast(const Stream&, Span<T8>, View<bool>); \
+    template void cast(const Stream&, Span<T8>, View<std::int32_t>); \
+    template void cast(const Stream&, Span<T8>, View<std::int64_t>); \
+    template void cast(const Stream&, Span<T8>, View<float>); \
+    template void cast(const Stream&, Span<T8>, View<std::uint8_t>); \
+    template void cast(const Stream&, Span<T8>, View<std::int8_t>); \
+    template void cast(const Stream&, Span<bool>, View<T8>); \
+    template void cast(const Stream&, Span<std::int32_t>, View<T8>); \
+    template void cast(const Stream&, Span<std::int64_t>, View<T8>); \
+    template void cast(const Stream&, Span<float>, View<T8>);
+    CV_CUDA_CAST_8BIT(std::uint8_t)
+    CV_CUDA_CAST_8BIT(std::int8_t)
+#undef CV_CUDA_CAST_8BIT
+
 }}}} /* namespace cv::dnn::cuda4dnn::kernels */

modules/dnn/src/cuda/col2im.cu

@@ -0,0 +1,79 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+// Copyright (C) 2026, BigVision LLC, all rights reserved.
+// Third party copyrights are property of their respective owners.
+
+#include <cuda_runtime.h>
+#include <cuda_fp16.h>
+
+#include "types.hpp"
+#include "grid_stride_range.hpp"
+#include "execution.hpp"
+
+#include "../cuda4dnn/csl/stream.hpp"
+#include "../cuda4dnn/csl/span.hpp"
+#include "../cuda4dnn/kernels/col2im.hpp"
+
+#include <opencv2/core.hpp>
+
+using namespace cv::dnn::cuda4dnn::csl;
+using namespace cv::dnn::cuda4dnn::csl::device;
+
+namespace cv { namespace dnn { namespace cuda4dnn { namespace kernels {
+
+    namespace raw {
+        template <class T>
+        __global__ void col2im_2d(Span<T> output, View<T> cols, View<T> bias, Col2ImParams p, size_type count)
+        {
+            const index_type in_plane = p.in_h * p.in_w;
+            const index_type out_plane = p.out_h * p.out_w;
+            for (auto i : grid_stride_range(count)) {
+                const index_type ox = i % p.out_w;
+                const index_type oy = (i / p.out_w) % p.out_h;
+                const index_type c = (i / out_plane) % p.channels;
+                const index_type n = i / (out_plane * p.channels);
+
+                float acc = bias.size() ? static_cast<float>(bias[p.channel_offset + c]) : 0.f;
+                const index_type batch_base = n * p.channels * p.kernel_h * p.kernel_w * in_plane;
+                for (index_type ky = 0; ky < p.kernel_h; ky++) {
+                    // input row iy feeds oy when iy * stride == oy + pad - ky * dilation
+                    const index_type ty = oy + p.pad_h - ky * p.dil_h;
+                    if (ty < 0 || ty % p.stride_h != 0)
+                        continue;
+                    const index_type iy = ty / p.stride_h;
+                    if (iy >= p.in_h)
+                        continue;
+                    for (index_type kx = 0; kx < p.kernel_w; kx++) {
+                        const index_type tx = ox + p.pad_w - kx * p.dil_w;
+                        if (tx < 0 || tx % p.stride_w != 0)
+                            continue;
+                        const index_type ix = tx / p.stride_w;
+                        if (ix >= p.in_w)
+                            continue;
+                        const index_type row = (c * p.kernel_h + ky) * p.kernel_w + kx;
+                        acc += static_cast<float>(cols[batch_base + row * in_plane + iy * p.in_w + ix]);
+                    }
+                }
+                output[(n * p.total_channels + p.channel_offset + c) * out_plane + oy * p.out_w + ox] = static_cast<T>(acc);
+            }
+        }
+    }
+
+    template <class T>
+    void col2im_2d(const Stream& stream, Span<T> output, View<T> cols, View<T> bias, const Col2ImParams& params)
+    {
+        const size_type count = params.batch * params.channels * params.out_h * params.out_w;
+        if (count == 0)
+            return;
+        auto kernel = raw::col2im_2d<T>;
+        auto policy = make_policy(kernel, count, 0, stream);
+        launch_kernel(kernel, policy, output, cols, bias, params, count);
+    }
+
+#if !defined(__CUDA_ARCH__) || (__CUDA_ARCH__ >= 530)
+    template void col2im_2d(const Stream&, Span<__half>, View<__half>, View<__half>, const Col2ImParams&);
+#endif
+    template void col2im_2d(const Stream&, Span<float>, View<float>, View<float>, const Col2ImParams&);
+
+}}}} /* namespace cv::dnn::cuda4dnn::kernels */

modules/dnn/src/cuda/cum_scan.cu

@@ -0,0 +1,108 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+// Copyright (C) 2026, BigVision LLC, all rights reserved.
+// Third party copyrights are property of their respective owners.
+#include <cuda_runtime.h>
+#include <cuda_fp16.h>
+
+#include "types.hpp"
+#include "grid_stride_range.hpp"
+#include "execution.hpp"
+
+#include "../cuda4dnn/csl/stream.hpp"
+#include "../cuda4dnn/csl/span.hpp"
+
+#include "../cuda4dnn/kernels/cum_scan.hpp"
+
+#include <opencv2/core.hpp>
+
+#include <cstddef>
+#include <cstdint>
+
+using namespace cv::dnn::cuda4dnn::csl;
+using namespace cv::dnn::cuda4dnn::csl::device;
+
+namespace cv { namespace dnn { namespace cuda4dnn { namespace kernels {
+
+    namespace raw {
+
+        /* fp16 accumulates in fp32; integral types must accumulate in themselves, since
+         * routing an int64 scan through float would silently lose precision. */
+        template <class T> struct work_type      { using type = T; };
+        template <> struct work_type<__half>     { using type = float; };
+
+        template <class T>
+        __global__ void cum_scan(Span<T> output, View<T> input,
+                                 std::int64_t outer_size, std::int64_t target_size, std::int64_t inner_size,
+                                 bool is_prod, bool exclusive, bool reverse)
+        {
+            using W = typename work_type<T>::type;
+
+            const T* src = input.data().get();
+            T* dst = output.data().get();
+
+            const std::int64_t outer_step_length = target_size * inner_size;
+
+            /* Same index arithmetic as the CPU scan; target_step is signed so that the
+             * reverse direction needs no separate code path. */
+            const std::int64_t target_start = reverse ? target_size - 1 : 0;
+            const std::int64_t target_stop  = reverse ? -1 : target_size;
+            const std::int64_t target_delta = reverse ? -1 : 1;
+            const std::int64_t target_step  = target_delta * inner_size;
+            const std::int64_t exclusive_delta = exclusive ? target_step : 0;
+
+            const std::int64_t nrows = outer_size * inner_size;
+
+            for (auto id : grid_stride_range(static_cast<device::size_type>(nrows))) {
+                const std::int64_t outer_idx = static_cast<std::int64_t>(id) / inner_size;
+                const std::int64_t inner_idx = static_cast<std::int64_t>(id) - outer_idx * inner_size;
+
+                const std::int64_t target_offset = outer_idx * outer_step_length;
+                const std::int64_t first = target_offset + target_start * inner_size + inner_idx;
+
+                // the running value stays in W, so fp16 really accumulates in fp32
+                W acc = exclusive ? (is_prod ? W(1) : W(0)) : static_cast<W>(src[first]);
+                dst[first] = static_cast<T>(acc);
+
+                for (std::int64_t t = target_start + target_delta; t != target_stop; t += target_delta) {
+                    const std::int64_t off = target_offset + t * inner_size + inner_idx;
+                    const W cur = static_cast<W>(src[off - exclusive_delta]);
+                    acc = static_cast<W>(is_prod ? acc * cur : acc + cur);
+                    dst[off] = static_cast<T>(acc);
+                }
+            }
+        }
+    }
+
+    template <class T>
+    void cum_scan(const Stream& stream,
+        Span<T> output, View<T> input,
+        std::size_t outer_size, std::size_t target_size, std::size_t inner_size,
+        bool is_prod, bool exclusive, bool reverse)
+    {
+        CV_Assert(outer_size > 0 && target_size > 0 && inner_size > 0);
+
+        const std::size_t nrows = outer_size * inner_size;
+        if (nrows == 0)
+            return;
+
+        auto kernel = raw::cum_scan<T>;
+        auto policy = make_policy(kernel, nrows, 0, stream);
+        launch_kernel(kernel, policy, output, input,
+                      static_cast<std::int64_t>(outer_size),
+                      static_cast<std::int64_t>(target_size),
+                      static_cast<std::int64_t>(inner_size),
+                      is_prod, exclusive, reverse);
+    }
+
+#if !defined(__CUDA_ARCH__) || (__CUDA_ARCH__ >= 530)
+    template void cum_scan(const Stream&, Span<__half>, View<__half>, std::size_t, std::size_t, std::size_t, bool, bool, bool);
+#endif
+    template void cum_scan(const Stream&, Span<float>, View<float>, std::size_t, std::size_t, std::size_t, bool, bool, bool);
+    template void cum_scan(const Stream&, Span<int32_t>, View<int32_t>, std::size_t, std::size_t, std::size_t, bool, bool, bool);
+    template void cum_scan(const Stream&, Span<int64_t>, View<int64_t>, std::size_t, std::size_t, std::size_t, bool, bool, bool);
+    template void cum_scan(const Stream&, Span<int8_t>, View<int8_t>, std::size_t, std::size_t, std::size_t, bool, bool, bool);
+    template void cum_scan(const Stream&, Span<uint8_t>, View<uint8_t>, std::size_t, std::size_t, std::size_t, bool, bool, bool);
+
+}}}} /* namespace cv::dnn::cuda4dnn::kernels */

modules/dnn/src/cuda/eltwise_ops.cu

@@ -355,6 +355,26 @@ void eltwise_pow_2(const Stream& stream, TensorSpan<T> output, TensorView<T> x,
     eltwise_op<T, PowFunctor<T>>(stream, output, x, y);
 }
 
+template <class T>
+void eltwise_prelu_2(const Stream& stream, TensorSpan<T> output, TensorView<T> x, TensorView<T> slope) {
+    eltwise_op<T, PReluFunctor<T>>(stream, output, x, slope);
+}
+
+template <class T>
+void eltwise_bitand_2(const Stream& stream, TensorSpan<T> output, TensorView<T> x, TensorView<T> y) {
+    eltwise_op<T, BitAndFunctor<T>>(stream, output, x, y);
+}
+
+template <class T>
+void eltwise_bitor_2(const Stream& stream, TensorSpan<T> output, TensorView<T> x, TensorView<T> y) {
+    eltwise_op<T, BitOrFunctor<T>>(stream, output, x, y);
+}
+
+template <class T>
+void eltwise_bitxor_2(const Stream& stream, TensorSpan<T> output, TensorView<T> x, TensorView<T> y) {
+    eltwise_op<T, BitXorFunctor<T>>(stream, output, x, y);
+}
+
 #if !defined(__CUDA_ARCH__) || (__CUDA_ARCH__ >= 530)
     template void eltwise_mod_2(const Stream& stream, TensorSpan<__half> output, TensorView<__half> x, TensorView<__half> y);
     template void eltwise_fmod_2(const Stream& stream, TensorSpan<__half> output, TensorView<__half> x, TensorView<__half> y);
@@ -422,5 +442,26 @@ void eltwise_pow_2(const Stream& stream, TensorSpan<T> output, TensorView<T> x,
     template void eltwise_min_2(const Stream& stream, TensorSpan<int64_t> output, TensorView<int64_t> x, TensorView<int64_t> y);
     template void eltwise_pow_2(const Stream& stream, TensorSpan<int64_t> output, TensorView<int64_t> x, TensorView<int64_t> y);
 
+#if !defined(__CUDA_ARCH__) || (__CUDA_ARCH__ >= 530)
+    template void eltwise_prelu_2(const Stream&, TensorSpan<__half>, TensorView<__half>, TensorView<__half>);
+#endif
+    template void eltwise_prelu_2(const Stream&, TensorSpan<float>, TensorView<float>, TensorView<float>);
+    template void eltwise_prelu_2(const Stream&, TensorSpan<int8_t>, TensorView<int8_t>, TensorView<int8_t>);
+    template void eltwise_prelu_2(const Stream&, TensorSpan<uint8_t>, TensorView<uint8_t>, TensorView<uint8_t>);
+    template void eltwise_prelu_2(const Stream&, TensorSpan<int32_t>, TensorView<int32_t>, TensorView<int32_t>);
+    template void eltwise_prelu_2(const Stream&, TensorSpan<int64_t>, TensorView<int64_t>, TensorView<int64_t>);
+
+    template void eltwise_bitand_2(const Stream&, TensorSpan<int8_t>, TensorView<int8_t>, TensorView<int8_t>);
+    template void eltwise_bitand_2(const Stream&, TensorSpan<uint8_t>, TensorView<uint8_t>, TensorView<uint8_t>);
+    template void eltwise_bitand_2(const Stream&, TensorSpan<int32_t>, TensorView<int32_t>, TensorView<int32_t>);
+    template void eltwise_bitand_2(const Stream&, TensorSpan<int64_t>, TensorView<int64_t>, TensorView<int64_t>);
+    template void eltwise_bitor_2(const Stream&, TensorSpan<int8_t>, TensorView<int8_t>, TensorView<int8_t>);
+    template void eltwise_bitor_2(const Stream&, TensorSpan<uint8_t>, TensorView<uint8_t>, TensorView<uint8_t>);
+    template void eltwise_bitor_2(const Stream&, TensorSpan<int32_t>, TensorView<int32_t>, TensorView<int32_t>);
+    template void eltwise_bitor_2(const Stream&, TensorSpan<int64_t>, TensorView<int64_t>, TensorView<int64_t>);
+    template void eltwise_bitxor_2(const Stream&, TensorSpan<int8_t>, TensorView<int8_t>, TensorView<int8_t>);
+    template void eltwise_bitxor_2(const Stream&, TensorSpan<uint8_t>, TensorView<uint8_t>, TensorView<uint8_t>);
+    template void eltwise_bitxor_2(const Stream&, TensorSpan<int32_t>, TensorView<int32_t>, TensorView<int32_t>);
+    template void eltwise_bitxor_2(const Stream&, TensorSpan<int64_t>, TensorView<int64_t>, TensorView<int64_t>);
 
 }}}} /* namespace cv::dnn::cuda4dnn::kernels */

modules/dnn/src/cuda/functors.hpp

@@ -9,6 +9,8 @@
 
 #include "math.hpp"
 
+#include <type_traits>
+
 #include "../cuda4dnn/csl/nvcc_defs.hpp"
 
 namespace cv { namespace dnn { namespace cuda4dnn { namespace kernels {
@@ -851,6 +853,63 @@ struct FModFunctor {
     }
 };
 
+// ONNX PRelu with a broadcast slope: y is the slope
+template <class T>
+struct PReluFunctor {
+    struct Params {
+        CUDA4DNN_HOST_DEVICE Params() {}
+    };
+
+    CUDA4DNN_DEVICE PReluFunctor() { }
+    CUDA4DNN_DEVICE PReluFunctor(const Params& params) { }
+
+    CUDA4DNN_DEVICE T operator()(T x, T y) {
+        return apply(x, y, std::is_unsigned<T>{});
+    }
+
+    // an unsigned value is never negative, so PRelu passes it through; an overload
+    // instead of if constexpr (C++17), which .cu files built as C++14 can't use
+    CUDA4DNN_DEVICE static T apply(T x, T, std::true_type) { return x; }
+    CUDA4DNN_DEVICE static T apply(T x, T y, std::false_type) { return x >= T(0) ? x : T(x * y); }
+};
+
+// bitwise functors are instantiated for integer types only
+template <class T>
+struct BitAndFunctor {
+    struct Params {
+        CUDA4DNN_HOST_DEVICE Params() {}
+    };
+
+    CUDA4DNN_DEVICE BitAndFunctor() { }
+    CUDA4DNN_DEVICE BitAndFunctor(const Params& params) { }
+
+    CUDA4DNN_DEVICE T operator()(T x, T y) { return x & y; }
+};
+
+template <class T>
+struct BitOrFunctor {
+    struct Params {
+        CUDA4DNN_HOST_DEVICE Params() {}
+    };
+
+    CUDA4DNN_DEVICE BitOrFunctor() { }
+    CUDA4DNN_DEVICE BitOrFunctor(const Params& params) { }
+
+    CUDA4DNN_DEVICE T operator()(T x, T y) { return x | y; }
+};
+
+template <class T>
+struct BitXorFunctor {
+    struct Params {
+        CUDA4DNN_HOST_DEVICE Params() {}
+    };
+
+    CUDA4DNN_DEVICE BitXorFunctor() { }
+    CUDA4DNN_DEVICE BitXorFunctor(const Params& params) { }
+
+    CUDA4DNN_DEVICE T operator()(T x, T y) { return x ^ y; }
+};
+
 template <class T>
 struct PowFunctor {
     struct Params {

modules/dnn/src/cuda/gather_elements.cu

@@ -0,0 +1,90 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+// Copyright (C) 2026, BigVision LLC, all rights reserved.
+// Third party copyrights are property of their respective owners.
+#include <cuda_runtime.h>
+#include <cuda_fp16.h>
+
+#include "types.hpp"
+#include "grid_stride_range.hpp"
+#include "execution.hpp"
+#include "fast_divmod.hpp"
+
+#include "../cuda4dnn/csl/stream.hpp"
+#include "../cuda4dnn/csl/span.hpp"
+
+#include "../cuda4dnn/kernels/gather_elements.hpp"
+
+#include <cstddef>
+#include <cstdint>
+
+using namespace cv::dnn::cuda4dnn::csl;
+using namespace cv::dnn::cuda4dnn::csl::device;
+
+namespace cv { namespace dnn { namespace cuda4dnn { namespace kernels {
+
+    namespace raw {
+        __host__ __device__ inline std::int64_t get_element_index_value(const void* index_data, std::size_t index_element_size, std::size_t offset) {
+            switch (index_element_size) {
+                case sizeof(std::int32_t):
+                    return *(reinterpret_cast<const std::int32_t*>(index_data) + offset);
+                case sizeof(std::int64_t):
+                    return *(reinterpret_cast<const std::int64_t*>(index_data) + offset);
+                default:
+                    break;
+            }
+            return 0;
+        }
+
+        template <class T>
+        __global__ void gather_elements(
+            Span<T> output, View<T> input,
+            const void* indices, std::size_t index_element_size,
+            std::int64_t axis_size_data, std::int64_t inner_size,
+            fast_divmod axis_inner_size, fast_divmod inner_size_divmod)
+        {
+            for (auto id : grid_stride_range(output.size())) {
+                int outer_index, axis_inner_offset;
+                axis_inner_size.divmod(id, outer_index, axis_inner_offset);
+                int axis_index, inner_index;
+                inner_size_divmod.divmod(axis_inner_offset, axis_index, inner_index);
+                (void)axis_index;
+
+                std::int64_t idx = get_element_index_value(indices, index_element_size, id);
+                idx = idx < 0 ? idx + axis_size_data : idx;
+                if (idx < 0 || idx >= axis_size_data) {
+                    output[id] = static_cast<T>(0);
+                    continue;
+                }
+
+                std::int64_t input_index = static_cast<std::int64_t>(outer_index) * axis_size_data * inner_size +
+                                           idx * inner_size + inner_index;
+                output[id] = input.data().get()[input_index];
+            }
+        }
+    }
+
+    template <class T>
+    void gather_elements(const Stream& stream,
+        Span<T> output, View<T> input,
+        const void* indices, std::size_t index_element_size,
+        std::int64_t axis_size_data, std::int64_t inner_size,
+        fast_divmod axis_inner_size, fast_divmod inner_size_divmod)
+    {
+        auto kernel = raw::gather_elements<T>;
+        auto policy = make_policy(kernel, output.size(), 0, stream);
+        launch_kernel(kernel, policy, output, input, indices, index_element_size,
+                      axis_size_data, inner_size, axis_inner_size, inner_size_divmod);
+    }
+
+#if !defined(__CUDA_ARCH__) || (__CUDA_ARCH__ >= 530)
+    template void gather_elements(const Stream&, Span<__half>, View<__half>, const void*, std::size_t, std::int64_t, std::int64_t, fast_divmod, fast_divmod);
+#endif
+    template void gather_elements(const Stream&, Span<float>, View<float>, const void*, std::size_t, std::int64_t, std::int64_t, fast_divmod, fast_divmod);
+    template void gather_elements(const Stream&, Span<int8_t>, View<int8_t>, const void*, std::size_t, std::int64_t, std::int64_t, fast_divmod, fast_divmod);
+    template void gather_elements(const Stream&, Span<uint8_t>, View<uint8_t>, const void*, std::size_t, std::int64_t, std::int64_t, fast_divmod, fast_divmod);
+    template void gather_elements(const Stream&, Span<int32_t>, View<int32_t>, const void*, std::size_t, std::int64_t, std::int64_t, fast_divmod, fast_divmod);
+    template void gather_elements(const Stream&, Span<int64_t>, View<int64_t>, const void*, std::size_t, std::int64_t, std::int64_t, fast_divmod, fast_divmod);
+
+}}}} /* namespace cv::dnn::cuda4dnn::kernels */

modules/dnn/src/cuda/gather_nd.cu

@@ -0,0 +1,118 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+// Copyright (C) 2026, BigVision LLC, all rights reserved.
+// Third party copyrights are property of their respective owners.
+#include <cuda_runtime.h>
+#include <cuda_fp16.h>
+
+#include "array.hpp"
+#include "types.hpp"
+#include "grid_stride_range.hpp"
+#include "execution.hpp"
+
+#include "../cuda4dnn/csl/stream.hpp"
+#include "../cuda4dnn/csl/span.hpp"
+#include "../cuda4dnn/csl/tensor.hpp"
+
+#include "../cuda4dnn/kernels/gather_nd.hpp"
+
+#include <opencv2/core.hpp>
+
+#include <cstddef>
+#include <cstdint>
+#include <vector>
+
+using namespace cv::dnn::cuda4dnn::csl;
+using namespace cv::dnn::cuda4dnn::csl::device;
+
+namespace cv { namespace dnn { namespace cuda4dnn { namespace kernels {
+
+    namespace raw {
+        template <class T, class TIdx>
+        __global__ void gather_nd(
+            Span<T> output, View<TIdx> indices, View<T> input,
+            std::int64_t last_index_dimension,
+            array<std::int64_t, CSL_MAX_TENSOR_RANK> element_counts,
+            array<std::int64_t, CSL_MAX_TENSOR_RANK> input_dims,
+            std::int64_t inner_size,
+            std::int64_t batch_stride,
+            std::int64_t num_slices_per_batch)
+        {
+            for (auto id : grid_stride_range(output.size())) {
+                const std::int64_t slice = static_cast<std::int64_t>(id) / inner_size;
+                const std::int64_t inner = static_cast<std::int64_t>(id) - slice * inner_size;
+
+                std::int64_t data_offset = 0;
+                const std::int64_t indices_start = last_index_dimension * slice;
+                for (std::int64_t j = 0; j < last_index_dimension; ++j) {
+                    std::int64_t index = static_cast<std::int64_t>(indices.data().get()[indices_start + j]);
+
+                    const std::int64_t dim_value = input_dims[j];
+                    if (index >= 0) {
+                        if (index >= dim_value)
+                            index = dim_value - 1;
+                    } else {
+                        if (index < -dim_value)
+                            index = 0;
+                        else
+                            index += dim_value;
+                    }
+
+                    data_offset += index * element_counts[j];
+                }
+
+                if (batch_stride > 0)
+                    data_offset += (slice / num_slices_per_batch) * batch_stride;
+
+                output[id] = input.data().get()[data_offset + inner];
+            }
+        }
+    }
+
+    template <class T, class TIdx>
+    void gather_nd(const Stream& stream,
+        Span<T> output, View<TIdx> indices, View<T> input,
+        std::int64_t last_index_dimension,
+        const std::vector<std::int64_t>& element_counts,
+        const std::vector<std::int64_t>& input_dims,
+        std::int64_t inner_size,
+        std::int64_t batch_stride,
+        std::int64_t num_slices_per_batch)
+    {
+        CV_Assert(last_index_dimension <= CSL_MAX_TENSOR_RANK);
+        CV_Assert(element_counts.size() == static_cast<std::size_t>(last_index_dimension));
+        CV_Assert(input_dims.size() == static_cast<std::size_t>(last_index_dimension));
+        CV_Assert(inner_size > 0);
+        CV_Assert(num_slices_per_batch > 0);
+
+        if (output.size() == 0)
+            return;
+
+        array<std::int64_t, CSL_MAX_TENSOR_RANK> element_counts_k, input_dims_k;
+        element_counts_k.assign(std::begin(element_counts), std::end(element_counts));
+        input_dims_k.assign(std::begin(input_dims), std::end(input_dims));
+
+        auto kernel = raw::gather_nd<T, TIdx>;
+        auto policy = make_policy(kernel, output.size(), 0, stream);
+        launch_kernel(kernel, policy, output, indices, input,
+                      last_index_dimension, element_counts_k, input_dims_k,
+                      inner_size, batch_stride, num_slices_per_batch);
+    }
+
+#define CV_GATHER_ND_INST(T) \
+    template void gather_nd(const Stream&, Span<T>, View<std::int32_t>, View<T>, std::int64_t, const std::vector<std::int64_t>&, const std::vector<std::int64_t>&, std::int64_t, std::int64_t, std::int64_t); \
+    template void gather_nd(const Stream&, Span<T>, View<std::int64_t>, View<T>, std::int64_t, const std::vector<std::int64_t>&, const std::vector<std::int64_t>&, std::int64_t, std::int64_t, std::int64_t);
+
+#if !defined(__CUDA_ARCH__) || (__CUDA_ARCH__ >= 530)
+    CV_GATHER_ND_INST(__half)
+#endif
+    CV_GATHER_ND_INST(float)
+    CV_GATHER_ND_INST(int8_t)
+    CV_GATHER_ND_INST(uint8_t)
+    CV_GATHER_ND_INST(int32_t)
+    CV_GATHER_ND_INST(int64_t)
+
+#undef CV_GATHER_ND_INST
+
+}}}} /* namespace cv::dnn::cuda4dnn::kernels */

modules/dnn/src/cuda/grid_sample.cu

@@ -0,0 +1,162 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+// Copyright (C) 2026, BigVision LLC, all rights reserved.
+// Third party copyrights are property of their respective owners.
+#include <cuda_runtime.h>
+#include <cuda_fp16.h>
+
+#include "types.hpp"
+#include "grid_stride_range.hpp"
+#include "execution.hpp"
+
+#include "../cuda4dnn/csl/stream.hpp"
+#include "../cuda4dnn/csl/span.hpp"
+
+#include "../cuda4dnn/kernels/grid_sample.hpp"
+
+#include <opencv2/core.hpp>
+
+#include <cstdint>
+
+using namespace cv::dnn::cuda4dnn::csl;
+using namespace cv::dnn::cuda4dnn::csl::device;
+
+namespace cv { namespace dnn { namespace cuda4dnn { namespace kernels {
+
+    namespace raw {
+
+        enum { M_NEAREST = 0, M_BILINEAR = 1, M_BICUBIC = 2 };
+        enum { P_ZEROS = 0, P_BORDER = 1, P_REFLECTION = 2 };
+
+        __device__ inline float reflect_coord(float x, int limit, bool align_corners) {
+            if (limit <= 1) return 0.f;
+
+            const float minv = align_corners ? 0.f : -0.5f;
+            const float maxv = align_corners ? float(limit - 1) : float(limit) - 0.5f;
+            const float m = maxv - minv;
+            const float two_m = 2.f * m;
+
+            float t = fmodf(x - minv, two_m);
+            if (t < 0) t += two_m;
+            if (t > m) t = two_m - t;
+            return t + minv;
+        }
+
+        __device__ inline void cubic_coeffs(float x, float A, float* c) {
+            c[0] = ((A * (x + 1.0f) - 5.0f * A) * (x + 1.0f) + 8.0f * A) * (x + 1.0f) - 4.0f * A;
+            c[1] = ((A + 2.0f) * x - (A + 3.0f)) * x * x + 1.0f;
+            c[2] = ((A + 2.0f) * (1.0f - x) - (A + 3.0f)) * (1.0f - x) * (1.0f - x) + 1.0f;
+            c[3] = 1.0f - c[0] - c[1] - c[2];
+        }
+
+        /* The padding mode is decided per layer, so this branch is warp-uniform. */
+        template <class T>
+        __device__ inline float fetch(const T* baseNC, int yy, int xx, const GridSampleParams& p) {
+            int px = xx, py = yy;
+            if (p.padding == P_BORDER) {
+                px = static_cast<int>(fminf(float(p.width - 1),  fmaxf(0.f, float(px))));
+                py = static_cast<int>(fminf(float(p.height - 1), fmaxf(0.f, float(py))));
+            } else if (p.padding == P_REFLECTION) {
+                px = static_cast<int>(floorf(reflect_coord(float(px), p.width,  p.align_corners) + 0.5f));
+                py = static_cast<int>(floorf(reflect_coord(float(py), p.height, p.align_corners) + 0.5f));
+            }
+            if (px < 0 || py < 0 || px >= p.width || py >= p.height) return 0.f;
+            return static_cast<float>(baseNC[static_cast<std::int64_t>(py) * p.width + px]);
+        }
+
+        template <class T>
+        __global__ void grid_sample(Span<T> output, View<T> input, View<float> grid, GridSampleParams p)
+        {
+            const float delta = p.align_corners ? 1.f : 0.f;
+            const float xscale = 0.5f * (p.width - delta);
+            const float yscale = 0.5f * (p.height - delta);
+            const float xdelta = 0.5f * (p.width - delta)  + 0.5f * (delta - 1.f);
+            const float ydelta = 0.5f * (p.height - delta) + 0.5f * (delta - 1.f);
+
+            for (auto id : grid_stride_range(output.size())) {
+                int rem = id;
+                const int w = rem % p.output_width;  rem /= p.output_width;
+                const int h = rem % p.output_height; rem /= p.output_height;
+                const int c = rem % p.channels;      rem /= p.channels;
+                const int n = rem;
+
+                const T* baseNC = input.data().get() +
+                                  (static_cast<std::int64_t>(n) * p.channels + c) *
+                                  static_cast<std::int64_t>(p.height) * p.width;
+
+                const std::int64_t g = ((static_cast<std::int64_t>(n) * p.output_height + h) *
+                                        p.output_width + w) * 2;
+                const float nx = grid.data().get()[g + 0];
+                const float ny = grid.data().get()[g + 1];
+
+                const float xf = nx * xscale + xdelta;
+                const float yf = ny * yscale + ydelta;
+
+                float outv = 0.f;
+                if (p.mode == M_NEAREST) {
+                    outv = fetch<T>(baseNC, __float2int_rn(yf), __float2int_rn(xf), p);
+                } else if (p.mode == M_BILINEAR) {
+                    const int x0 = static_cast<int>(floorf(xf));
+                    const int y0 = static_cast<int>(floorf(yf));
+                    const float dx = xf - x0, dy = yf - y0;
+
+                    const float v00 = fetch<T>(baseNC, y0,     x0,     p);
+                    const float v01 = fetch<T>(baseNC, y0,     x0 + 1, p);
+                    const float v10 = fetch<T>(baseNC, y0 + 1, x0,     p);
+                    const float v11 = fetch<T>(baseNC, y0 + 1, x0 + 1, p);
+
+                    const float vx0 = v00 * (1.f - dx) + v01 * dx;
+                    const float vx1 = v10 * (1.f - dx) + v11 * dx;
+                    outv = vx0 * (1.f - dy) + vx1 * dy;
+                } else {
+                    const int x1 = static_cast<int>(floorf(xf));
+                    const int y1 = static_cast<int>(floorf(yf));
+                    const float tx = xf - x1, ty = yf - y1;
+
+                    float wx[4], wy[4];
+                    cubic_coeffs(tx, p.cubic_alpha, wx);
+                    cubic_coeffs(ty, p.cubic_alpha, wy);
+
+                    outv = 0.f;
+                    for (int j = 0; j < 4; ++j) {
+                        float rowv = 0.f;
+                        for (int i = 0; i < 4; ++i)
+                            rowv += fetch<T>(baseNC, y1 - 1 + j, x1 - 1 + i, p) * wx[i];
+                        outv += rowv * wy[j];
+                    }
+                }
+
+                output[id] = static_cast<T>(outv);
+            }
+        }
+    }
+
+    template <class T>
+    void grid_sample(const Stream& stream,
+        Span<T> output, View<T> input, View<float> grid,
+        const GridSampleParams& params)
+    {
+        CV_Assert(params.channels > 0 && params.height > 0 && params.width > 0);
+        CV_Assert(params.output_height > 0 && params.output_width > 0);
+        CV_Assert(params.mode >= 0 && params.mode <= 2);
+        CV_Assert(params.padding >= 0 && params.padding <= 2);
+
+        if (output.size() == 0)
+            return;
+
+        auto kernel = raw::grid_sample<T>;
+        auto policy = make_policy(kernel, output.size(), 0, stream);
+        launch_kernel(kernel, policy, output, input, grid, params);
+    }
+
+#if !defined(__CUDA_ARCH__) || (__CUDA_ARCH__ >= 530)
+    template void grid_sample(const Stream&, Span<__half>, View<__half>, View<float>, const GridSampleParams&);
+#endif
+    template void grid_sample(const Stream&, Span<float>, View<float>, View<float>, const GridSampleParams&);
+    template void grid_sample(const Stream&, Span<int8_t>, View<int8_t>, View<float>, const GridSampleParams&);
+    template void grid_sample(const Stream&, Span<uint8_t>, View<uint8_t>, View<float>, const GridSampleParams&);
+    template void grid_sample(const Stream&, Span<int32_t>, View<int32_t>, View<float>, const GridSampleParams&);
+    template void grid_sample(const Stream&, Span<int64_t>, View<int64_t>, View<float>, const GridSampleParams&);
+
+}}}} /* namespace cv::dnn::cuda4dnn::kernels */

modules/dnn/src/cuda/mvn.cu

@@ -34,21 +34,11 @@ namespace raw {
 
     constexpr int MEAN_SQR_SUM_BLOCK_SIZE = 256;
 
-    // one block per row: a warp-shuffle and shared-memory reduction, no per-element atomics
-    template <class T>
-    __global__ void reduce_mean_sqr_sum(Span<float> means, Span<float> sum_sqrs, View<T> input, size_type inner_size) {
+    // sums sum and sqr over the block with warp shuffles and shared memory; the totals land in thread 0
+    __device__ void block_sum2(float& sum, float& sqr) {
         __shared__ float warp_sums[MEAN_SQR_SUM_BLOCK_SIZE / GPU_WARP_SIZE];
         __shared__ float warp_sqrs[MEAN_SQR_SUM_BLOCK_SIZE / GPU_WARP_SIZE];
 
-        const index_type outer_idx = blockIdx.x;
-        const index_type row_start = outer_idx * inner_size;
-        float sum = 0.f, sqr = 0.f;
-        for (index_type i = threadIdx.x; i < inner_size; i += blockDim.x) {
-            const auto x = static_cast<float>(input[row_start + i]);
-            sum += x;
-            sqr += x * x;
-        }
-
         for (int offset = GPU_WARP_SIZE / 2; offset > 0; offset /= 2) {
             sum += WARP_SHFL_XOR(sum, offset);
             sqr += WARP_SHFL_XOR(sqr, offset);
@@ -68,10 +58,59 @@ namespace raw {
                 sum += WARP_SHFL_XOR(sum, offset);
                 sqr += WARP_SHFL_XOR(sqr, offset);
             }
-            if (lane == 0) {
-                means[outer_idx] = sum / inner_size;
-                sum_sqrs[outer_idx] = sqr;
+        }
+    }
+
+    // one block per row: a warp-shuffle and shared-memory reduction, no per-element atomics
+    template <class T>
+    __global__ void reduce_mean_sqr_sum(Span<float> means, Span<float> sum_sqrs, View<T> input, size_type inner_size) {
+        const index_type outer_idx = blockIdx.x;
+        const index_type row_start = outer_idx * inner_size;
+        float sum = 0.f, sqr = 0.f;
+        for (index_type i = threadIdx.x; i < inner_size; i += blockDim.x) {
+            const auto x = static_cast<float>(input[row_start + i]);
+            sum += x;
+            sqr += x * x;
+        }
+        block_sum2(sum, sqr);
+        if (threadIdx.x == 0) {
+            means[outer_idx] = sum / inner_size;
+            sum_sqrs[outer_idx] = sqr;
+        }
+    }
+
+    // a long row split over `splits` blocks: each writes the partial sums of its chunk
+    template <class T>
+    __global__ void reduce_mean_sqr_sum_partial(Span<float> part_sums, Span<float> part_sqrs, View<T> input,
+                                                size_type inner_size, size_type chunk, size_type splits) {
+        const index_type row = blockIdx.x / splits, split = blockIdx.x % splits;
+        const index_type row_start = row * inner_size;
+        const index_type begin = split * chunk;
+        const index_type end = begin + chunk < inner_size ? begin + chunk : inner_size;
+        float sum = 0.f, sqr = 0.f;
+        for (index_type i = begin + threadIdx.x; i < end; i += blockDim.x) {
+            const auto x = static_cast<float>(input[row_start + i]);
+            sum += x;
+            sqr += x * x;
+        }
+        block_sum2(sum, sqr);
+        if (threadIdx.x == 0) {
+            part_sums[row * splits + split] = sum;
+            part_sqrs[row * splits + split] = sqr;
+        }
+    }
+
+    // one thread per row adds the partials in split order, so the result does not depend on scheduling
+    __global__ void reduce_mean_sqr_sum_combine(Span<float> means, Span<float> sum_sqrs, View<float> part_sums,
+                                                View<float> part_sqrs, size_type splits, size_type inner_size) {
+        for (auto row : grid_stride_range(means.size())) {
+            float sum = 0.f, sqr = 0.f;
+            for (index_type s = 0; s < splits; s++) {
+                sum += part_sums[row * splits + s];
+                sqr += part_sqrs[row * splits + s];
             }
+            means[row] = sum / inner_size;
+            sum_sqrs[row] = sqr;
         }
     }
 
@@ -181,6 +220,56 @@ template void reduce_mean_sqr_sum(const Stream&, Span<float>, Span<float>, View<
 #endif
 template void reduce_mean_sqr_sum(const Stream&, Span<float>, Span<float>, View<float>, std::size_t);
 
+namespace {
+    // 64 loads per thread: enough work per block to hide memory latency, small enough to fill the GPU
+    constexpr std::size_t MEAN_SQR_SUM_CHUNK = raw::MEAN_SQR_SUM_BLOCK_SIZE * 64;
+
+    // a row gains from splitting only once it fills two chunks; short rows (LayerNorm) keep one block
+    std::size_t meanSqrSumSplits(std::size_t inner_size)
+    {
+        if (inner_size < 2 * MEAN_SQR_SUM_CHUNK)
+            return 1;
+        return (inner_size + MEAN_SQR_SUM_CHUNK - 1) / MEAN_SQR_SUM_CHUNK;
+    }
+}
+
+std::size_t reduce_mean_sqr_sum_scratch(std::size_t rows, std::size_t inner_size)
+{
+    const std::size_t splits = meanSqrSumSplits(inner_size);
+    return splits > 1 ? 2 * rows * splits : 0;
+}
+
+template <class T>
+void reduce_mean_sqr_sum(const Stream& stream, Span<float> means, Span<float> sum_sqrs, View<T> input,
+                         std::size_t inner_size, Span<float> scratch)
+{
+    const std::size_t rows = means.size();
+    const std::size_t splits = meanSqrSumSplits(inner_size);
+    if (splits == 1 || scratch.size() < 2 * rows * splits) {
+        reduce_mean_sqr_sum<T>(stream, means, sum_sqrs, input, inner_size);
+        return;
+    }
+    CV_Assert(input.size() / inner_size == rows && sum_sqrs.size() == rows);
+
+    const std::size_t chunk = (inner_size + splits - 1) / splits;
+    Span<float> part_sums(scratch.data(), rows * splits);
+    Span<float> part_sqrs(scratch.data() + rows * splits, rows * splits);
+
+    auto partial = raw::reduce_mean_sqr_sum_partial<T>;
+    launch_kernel(partial, execution_policy(dim3(static_cast<unsigned int>(rows * splits)),
+                                            dim3(raw::MEAN_SQR_SUM_BLOCK_SIZE), 0, stream),
+                  part_sums, part_sqrs, input, inner_size, chunk, splits);
+
+    auto combine = raw::reduce_mean_sqr_sum_combine;
+    auto policy = make_policy(combine, rows, 0, stream);
+    launch_kernel(combine, policy, means, sum_sqrs, View<float>(part_sums), View<float>(part_sqrs), splits, inner_size);
+}
+
+#if !defined(__CUDA_ARCH__) || (__CUDA_ARCH__ >= 530)
+template void reduce_mean_sqr_sum(const Stream&, Span<float>, Span<float>, View<__half>, std::size_t, Span<float>);
+#endif
+template void reduce_mean_sqr_sum(const Stream&, Span<float>, Span<float>, View<float>, std::size_t, Span<float>);
+
 void compute_normalization_scale(const Stream& stream, Span<float> scale, View<float> means, View<float> sum_sqrs, std::size_t inner_size, float eps)
 {
     CV_Assert(scale.size() == means.size());

modules/dnn/src/cuda/normalize.cu

@@ -41,8 +41,9 @@ namespace cv { namespace dnn { namespace cuda4dnn { namespace kernels {
 
         template <class T>
         __global__ void reciprocal(Span<T> output, T epsilon) {
+            // a zero slice scales to 0; in half, eps underflows to 0 and 1/0 would give 0 * inf = NaN
             for (auto idx : grid_stride_range(output.size()))
-                output[idx] = T(1) / (output[idx] + epsilon);
+                output[idx] = output[idx] == T(0) ? T(0) : T(1) / (output[idx] + epsilon);
         }
 
         template <class T>
@@ -60,7 +61,8 @@ namespace cv { namespace dnn { namespace cuda4dnn { namespace kernels {
         __global__ void rsqrt(Span<T> output, T epsilon) {
             for (auto idx : grid_stride_range(output.size())) {
                 using device::sqrt;
-                output[idx] = T(1) / sqrt(output[idx] + epsilon);
+                // a zero slice scales to 0; in half, eps underflows to 0 and 1/0 would give 0 * inf = NaN
+                output[idx] = output[idx] == T(0) ? T(0) : T(1) / sqrt(output[idx] + epsilon);
             }
         }
 
@@ -115,9 +117,51 @@ namespace cv { namespace dnn { namespace cuda4dnn { namespace kernels {
         launch_kernel(scale_kernel, policy, output, input, mid_size * inner_size, inner_size, sums);
     }
 
+    namespace raw {
+        template <class T>
+        __global__ void lrn_across_channels(Span<T> output, View<T> input, size_type channels, size_type plane_size,
+                                            int radius, float alpha_scaled, float bias, float beta)
+        {
+            for (auto i : grid_stride_range(output.size())) {
+                const index_type p = i % plane_size;
+                const index_type c = (i / plane_size) % channels;
+                const index_type n = i / (plane_size * channels);
+                const index_type c0 = c - radius < 0 ? 0 : c - radius;
+                const index_type c1 = c + radius >= channels ? channels - 1 : c + radius;
+
+                // same window as the CPU ChannelLRN, accumulated in float
+                float s = 0.f;
+                for (index_type k = c0; k <= c1; k++) {
+                    const float v = static_cast<float>(input[(n * channels + k) * plane_size + p]);
+                    s += v * v;
+                }
+                const float x = static_cast<float>(input[i]);
+                output[i] = static_cast<T>(x * powf(bias + alpha_scaled * s, -beta));
+            }
+        }
+    }
+
+    template <class T>
+    void lrn_across_channels(
+        const Stream& stream,
+        Span<T> output, View<T> input,
+        std::size_t batch, std::size_t channels, std::size_t plane_size,
+        int radius, float alpha_scaled, float bias, float beta)
+    {
+        CV_Assert(output.size() == input.size());
+        CV_Assert(output.size() == batch * channels * plane_size);
+        if (output.size() == 0)
+            return;
+        auto kernel = raw::lrn_across_channels<T>;
+        auto policy = make_policy(kernel, output.size(), 0, stream);
+        launch_kernel(kernel, policy, output, input, channels, plane_size, radius, alpha_scaled, bias, beta);
+    }
+
 #if !defined(__CUDA_ARCH__) || (__CUDA_ARCH__ >= 530)
     template void normalize(const Stream&, Span<__half>, View<__half>, std::size_t, std::size_t, std::size_t, std::size_t, __half, Span<__half>);
+    template void lrn_across_channels(const Stream&, Span<__half>, View<__half>, std::size_t, std::size_t, std::size_t, int, float, float, float);
 #endif
+    template void lrn_across_channels(const Stream&, Span<float>, View<float>, std::size_t, std::size_t, std::size_t, int, float, float, float);
     template void normalize(const Stream&, Span<float>, View<float>, std::size_t, std::size_t, std::size_t, std::size_t, float, Span<float>);
 
 }}}} /* namespace cv::dnn::cuda4dnn::kernels */

modules/dnn/src/cuda/one_hot.cu

@@ -0,0 +1,78 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+// Copyright (C) 2026, BigVision LLC, all rights reserved.
+// Third party copyrights are property of their respective owners.
+#include <cuda_runtime.h>
+#include <cuda_fp16.h>
+
+#include "types.hpp"
+#include "grid_stride_range.hpp"
+#include "execution.hpp"
+
+#include "../cuda4dnn/csl/stream.hpp"
+#include "../cuda4dnn/csl/span.hpp"
+
+#include "../cuda4dnn/kernels/one_hot.hpp"
+
+#include <opencv2/core.hpp>
+
+#include <cstddef>
+#include <cstdint>
+
+using namespace cv::dnn::cuda4dnn::csl;
+using namespace cv::dnn::cuda4dnn::csl::device;
+
+namespace cv { namespace dnn { namespace cuda4dnn { namespace kernels {
+
+    namespace raw {
+        template <class T, class TIdx>
+        __global__ void one_hot(Span<T> output, View<TIdx> indices, std::int64_t depth, std::int64_t inner,
+                                T off_value, T on_value)
+        {
+            for (auto id : grid_stride_range(output.size())) {
+                const std::int64_t i = id;
+                const std::int64_t lo = i % inner;
+                const std::int64_t outer = i / inner;
+                const std::int64_t d = outer % depth;
+                const std::int64_t hi = outer / depth;
+                // only [-depth, depth-1] is valid; anything else matches no d and stays all off_value
+                std::int64_t v = static_cast<std::int64_t>(indices.data().get()[hi * inner + lo]);
+                if (v < 0)
+                    v += depth;
+                output[id] = (v == d) ? on_value : off_value;
+            }
+        }
+    }
+
+    template <class T, class TIdx>
+    void one_hot(const Stream& stream, Span<T> output, View<TIdx> indices,
+                 std::size_t depth, std::size_t inner, T off_value, T on_value)
+    {
+        CV_Assert(depth > 0 && inner > 0);
+        CV_Assert(output.size() == indices.size() * depth);
+        if (output.size() == 0)
+            return;
+
+        auto kernel = raw::one_hot<T, TIdx>;
+        auto policy = make_policy(kernel, output.size(), 0, stream);
+        launch_kernel(kernel, policy, output, indices, static_cast<std::int64_t>(depth),
+                      static_cast<std::int64_t>(inner), off_value, on_value);
+    }
+
+#if !defined(__CUDA_ARCH__) || (__CUDA_ARCH__ >= 530)
+    template void one_hot(const Stream&, Span<__half>, View<int32_t>, std::size_t, std::size_t, __half, __half);
+    template void one_hot(const Stream&, Span<__half>, View<int64_t>, std::size_t, std::size_t, __half, __half);
+#endif
+    template void one_hot(const Stream&, Span<float>, View<int32_t>, std::size_t, std::size_t, float, float);
+    template void one_hot(const Stream&, Span<float>, View<int64_t>, std::size_t, std::size_t, float, float);
+    template void one_hot(const Stream&, Span<int8_t>, View<int32_t>, std::size_t, std::size_t, int8_t, int8_t);
+    template void one_hot(const Stream&, Span<int8_t>, View<int64_t>, std::size_t, std::size_t, int8_t, int8_t);
+    template void one_hot(const Stream&, Span<uint8_t>, View<int32_t>, std::size_t, std::size_t, uint8_t, uint8_t);
+    template void one_hot(const Stream&, Span<uint8_t>, View<int64_t>, std::size_t, std::size_t, uint8_t, uint8_t);
+    template void one_hot(const Stream&, Span<int32_t>, View<int32_t>, std::size_t, std::size_t, int32_t, int32_t);
+    template void one_hot(const Stream&, Span<int32_t>, View<int64_t>, std::size_t, std::size_t, int32_t, int32_t);
+    template void one_hot(const Stream&, Span<int64_t>, View<int32_t>, std::size_t, std::size_t, int64_t, int64_t);
+    template void one_hot(const Stream&, Span<int64_t>, View<int64_t>, std::size_t, std::size_t, int64_t, int64_t);
+
+}}}} /* namespace cv::dnn::cuda4dnn::kernels */

modules/dnn/src/cuda/padding.cu

@@ -16,6 +16,9 @@
 #include "../cuda4dnn/csl/tensor.hpp"
 #include "../cuda4dnn/csl/span.hpp"
 
+#include "../cuda4dnn/kernels/fill_copy.hpp"
+#include "../cuda4dnn/kernels/padding.hpp"
+
 #include <opencv2/core.hpp>
 
 #include <cstddef>
@@ -64,8 +67,114 @@ namespace cv { namespace dnn { namespace cuda4dnn { namespace kernels {
                 output[i] = input[iidx];
             }
         }
+
+        template <class T, std::size_t Rank>
+        __global__ void pad_nd(
+            Span<T> output, array<size_type, Rank> out_strides,
+            View<T> input, array<size_type, Rank> in_strides, array<index_type, Rank> in_shape,
+            array<index_type, Rank> pads_begin, int mode, T value)
+        {
+            for (auto i : grid_stride_range(output.size())) {
+                index_type iidx = 0;
+                bool outside = false;
+                for (int j = 0; j < Rank; j++) {
+                    const index_type n = in_shape[j];
+                    const index_type out_j = j == 0 ? i / out_strides[0] : (i % out_strides[j - 1]) / out_strides[j];
+                    index_type p = out_j - pads_begin[j];
+                    if (p < 0 || p >= n) {
+                        if (mode == static_cast<int>(kernels::PadMode::CONSTANT)) {
+                            outside = true;
+                            break;
+                        } else if (mode == static_cast<int>(kernels::PadMode::REPLICATE)) {
+                            p = p < 0 ? 0 : n - 1;
+                        } else if (mode == static_cast<int>(kernels::PadMode::WRAP)) {
+                            p = ((p % n) + n) % n;
+                        } else if (n == 1) {
+                            p = 0;
+                        } else {
+                            // reflect101 may bounce more than once when a pad exceeds the axis
+                            while (p < 0 || p >= n)
+                                p = p < 0 ? -p : 2 * n - 2 - p;
+                        }
+                    }
+                    iidx += p * in_strides[j];
+                }
+                output[i] = outside ? value : input[iidx];
+            }
+        }
     }
 
+    template <class T, std::size_t Rank> static
+    void launch_pad_nd(
+        const Stream& stream,
+        Span<T> output, const std::vector<std::size_t>& outStride,
+        View<T> input, const std::vector<std::size_t>& inStride, const std::vector<std::size_t>& inShape,
+        const std::vector<int>& padsBegin, int mode, T value)
+    {
+        CV_Assert(outStride.size() == Rank && inStride.size() == Rank);
+        CV_Assert(inShape.size() == Rank && padsBegin.size() == Rank);
+
+        array<size_type, Rank> outStride_k, inStride_k;
+        outStride_k.assign(std::begin(outStride), std::end(outStride));
+        inStride_k.assign(std::begin(inStride), std::end(inStride));
+
+        array<index_type, Rank> inShape_k, padsBegin_k;
+        for (int i = 0; i < Rank; i++) {
+            inShape_k[i] = inShape[i];
+            padsBegin_k[i] = padsBegin[i];
+        }
+
+        auto kernel = raw::pad_nd<T, Rank>;
+        auto policy = make_policy(kernel, output.size(), 0, stream);
+        launch_kernel(kernel, policy, output, outStride_k, input, inStride_k, inShape_k, padsBegin_k, mode, value);
+    }
+
+    GENERATE_KERNEL_DISPATCHER(pad_nd_dispatcher, launch_pad_nd);
+
+    template <class T>
+    void pad_nd(
+        const Stream& stream,
+        TensorSpan<T> output, TensorView<T> input,
+        const std::vector<int>& pads_begin, PadMode mode, T value)
+    {
+        CV_Assert(output.rank() == input.rank());
+        CV_Assert(output.rank() == pads_begin.size());
+        if (output.size() == 0)
+            return;
+        // an empty input leaves only constant padding
+        if (input.size() == 0) {
+            CV_Assert(mode == PadMode::CONSTANT);
+            fill<T>(stream, output, value);
+            return;
+        }
+
+        const auto rank = input.rank();
+        auto inShape = input.shape_as_vector();
+        auto outShape = output.shape_as_vector();
+
+        std::vector<std::size_t> inStride(rank), outStride(rank);
+        inStride.back() = 1;
+        outStride.back() = 1;
+        for (int i = (int)rank - 2; i >= 0; i--) {
+            inStride[i] = inStride[i + 1] * inShape[i + 1];
+            outStride[i] = outStride[i + 1] * outShape[i + 1];
+        }
+
+        CV_Assert(1 <= rank && rank <= CSL_MAX_TENSOR_RANK);
+        pad_nd_dispatcher<T, 1, CSL_MAX_TENSOR_RANK>(rank, stream, output, outStride, input, inStride, inShape,
+                                                     pads_begin, static_cast<int>(mode), value);
+    }
+
+#if !defined(__CUDA_ARCH__) || (__CUDA_ARCH__ >= 530)
+    template void pad_nd(const Stream&, TensorSpan<__half>, TensorView<__half>, const std::vector<int>&, PadMode, __half);
+#endif
+    template void pad_nd(const Stream&, TensorSpan<float>, TensorView<float>, const std::vector<int>&, PadMode, float);
+    template void pad_nd(const Stream&, TensorSpan<int8_t>, TensorView<int8_t>, const std::vector<int>&, PadMode, int8_t);
+    template void pad_nd(const Stream&, TensorSpan<uint8_t>, TensorView<uint8_t>, const std::vector<int>&, PadMode, uint8_t);
+    template void pad_nd(const Stream&, TensorSpan<int32_t>, TensorView<int32_t>, const std::vector<int>&, PadMode, int32_t);
+    template void pad_nd(const Stream&, TensorSpan<int64_t>, TensorView<int64_t>, const std::vector<int>&, PadMode, int64_t);
+    template void pad_nd(const Stream&, TensorSpan<bool>, TensorView<bool>, const std::vector<int>&, PadMode, bool);
+
     template <class T, std::size_t Rank> static
     void launch_copy_with_reflection101(
         const Stream& stream,

modules/dnn/src/cuda/quantize_dequantize.cu

@@ -0,0 +1,177 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+
+#include <cuda_runtime.h>
+
+#include <cub/block/block_reduce.cuh>
+
+#include "grid_stride_range.hpp"
+#include "execution.hpp"
+
+#include "../cuda4dnn/csl/stream.hpp"
+#include "../cuda4dnn/csl/span.hpp"
+
+#include "../cuda4dnn/kernels/quantize_dequantize.hpp"
+
+#include <opencv2/core.hpp>
+
+#include <algorithm>
+#include <cstddef>
+#include <cstdint>
+
+using namespace cv::dnn::cuda4dnn::csl;
+using namespace cv::dnn::cuda4dnn::csl::device;
+
+namespace cv { namespace dnn { namespace cuda4dnn { namespace kernels {
+
+    namespace raw {
+
+        template <class T>
+        __global__ void dequantize_linear(const T* input, const float* scale, const T* zero_point,
+                                           float* output, std::size_t sz_a, std::size_t slice_size, std::size_t n)
+        {
+            for (auto i : grid_stride_range(n)) {
+                std::size_t a_idx = (i / slice_size) % sz_a;
+                T zp = zero_point ? zero_point[a_idx] : T(0);
+                output[i] = (static_cast<float>(input[i]) - static_cast<float>(zp)) * scale[a_idx];
+            }
+        }
+
+        template <class T> __device__ __forceinline__ T saturate_quantized(int v);
+        template <> __device__ __forceinline__ std::int8_t saturate_quantized<std::int8_t>(int v) {
+            return static_cast<std::int8_t>(::min(::max(v, -128), 127));
+        }
+        template <> __device__ __forceinline__ std::uint8_t saturate_quantized<std::uint8_t>(int v) {
+            return static_cast<std::uint8_t>(::min(::max(v, 0), 255));
+        }
+
+        template <class T>
+        __global__ void quantize_linear(const float* input, const float* scale, const T* zero_point,
+                                         T* output, std::size_t sz_a, std::size_t slice_size, std::size_t n)
+        {
+            for (auto i : grid_stride_range(n)) {
+                std::size_t a_idx = (i / slice_size) % sz_a;
+                int zp = zero_point ? static_cast<int>(zero_point[a_idx]) : 0;
+                // clamp before converting: __float2int_rn turns +Inf or a huge value into
+                // INT_MIN, which would saturate to qmin instead of qmax
+                float scaled = input[i] / scale[a_idx];
+                scaled = (scaled != scaled) ? 0.f : fminf(fmaxf(scaled, -1024.f), 1024.f);
+                int rounded = __float2int_rn(scaled);
+                output[i] = saturate_quantized<T>(rounded + zp);
+            }
+        }
+
+        // 256 threads, as the module's other block reductions (rms_norm.cu, topk.cu, reduce.cu, mvn.cu)
+        constexpr int DYNAMIC_QUANTIZE_BLOCK = 256;
+
+        struct MinOp { __device__ float operator()(float a, float b) const { return fminf(a, b); } };
+        struct MaxOp { __device__ float operator()(float a, float b) const { return fmaxf(a, b); } };
+
+        // one (min, max) pair per block; zero starts both, since the range must include it
+        __global__ void dynamic_quantize_minmax(const float* input, size_type n, float* pmin, float* pmax)
+        {
+            using BlockReduce = cub::BlockReduce<float, DYNAMIC_QUANTIZE_BLOCK>;
+            __shared__ typename BlockReduce::TempStorage storage;
+
+            float lo = 0.f, hi = 0.f;
+            for (size_type i = blockIdx.x * DYNAMIC_QUANTIZE_BLOCK + threadIdx.x; i < n; i += gridDim.x * DYNAMIC_QUANTIZE_BLOCK) {
+                lo = fminf(lo, input[i]);
+                hi = fmaxf(hi, input[i]);
+            }
+            lo = BlockReduce(storage).Reduce(lo, MinOp());
+            __syncthreads();
+            hi = BlockReduce(storage).Reduce(hi, MaxOp());
+            if (threadIdx.x == 0) {
+                pmin[blockIdx.x] = lo;
+                pmax[blockIdx.x] = hi;
+            }
+        }
+
+        // folds the partials and derives the scale and zero point exactly as the CPU layer does
+        __global__ void dynamic_quantize_params(const float* pmin, const float* pmax, int nparts,
+                                                float* scale, std::uint8_t* zero_point)
+        {
+            using BlockReduce = cub::BlockReduce<float, DYNAMIC_QUANTIZE_BLOCK>;
+            __shared__ typename BlockReduce::TempStorage storage;
+
+            float lo = 0.f, hi = 0.f;
+            for (int i = threadIdx.x; i < nparts; i += DYNAMIC_QUANTIZE_BLOCK) {
+                lo = fminf(lo, pmin[i]);
+                hi = fmaxf(hi, pmax[i]);
+            }
+            lo = BlockReduce(storage).Reduce(lo, MinOp());
+            __syncthreads();
+            hi = BlockReduce(storage).Reduce(hi, MaxOp());
+            if (threadIdx.x == 0) {
+                const float qmin = 0.f, qmax = 255.f;
+                float y_scale = (hi - lo) / (qmax - qmin);
+                if (y_scale == 0.f)
+                    y_scale = 1.f;
+                const int zp = __float2int_rn(qmin - lo / y_scale);
+                scale[0] = y_scale;
+                zero_point[0] = saturate_quantized<std::uint8_t>(zp);
+            }
+        }
+
+    }
+
+    template <class T>
+    void dequantize_linear(const Stream& stream, Span<float> output, View<T> input,
+                            View<float> scale, View<T> zero_point,
+                            std::size_t sz_a, std::size_t slice_size)
+    {
+        CV_Assert(output.size() == input.size());
+        CV_Assert(zero_point.empty() || zero_point.size() == scale.size());
+
+        auto kernel = raw::dequantize_linear<T>;
+        auto policy = make_policy(kernel, output.size(), 0, stream);
+        launch_kernel(kernel, policy, input.data().get(), scale.data().get(),
+                      zero_point.empty() ? nullptr : zero_point.data().get(),
+                      output.data().get(), sz_a, slice_size, output.size());
+    }
+
+    template <class T>
+    void quantize_linear(const Stream& stream, Span<T> output, View<float> input,
+                          View<float> scale, View<T> zero_point,
+                          std::size_t sz_a, std::size_t slice_size)
+    {
+        CV_Assert(output.size() == input.size());
+        CV_Assert(zero_point.empty() || zero_point.size() == scale.size());
+
+        auto kernel = raw::quantize_linear<T>;
+        auto policy = make_policy(kernel, output.size(), 0, stream);
+        launch_kernel(kernel, policy, input.data().get(), scale.data().get(),
+                      zero_point.empty() ? nullptr : zero_point.data().get(),
+                      output.data().get(), sz_a, slice_size, output.size());
+    }
+
+    std::size_t dynamic_quantize_partials(std::size_t n)
+    {
+        const std::size_t block = raw::DYNAMIC_QUANTIZE_BLOCK;
+        return std::max<std::size_t>(1, std::min<std::size_t>((n + block - 1) / block, block));
+    }
+
+    void dynamic_quantize_scale(const Stream& stream, Span<float> scale, Span<std::uint8_t> zero_point,
+                                View<float> input, Span<float> scratch)
+    {
+        CV_Assert(scale.size() == 1 && zero_point.size() == 1);
+        const std::size_t nparts = dynamic_quantize_partials(input.size());
+        CV_Assert(scratch.size() >= 2 * nparts);
+        float* pmin = scratch.data().get();
+        float* pmax = pmin + nparts;
+
+        auto minmax = raw::dynamic_quantize_minmax;
+        launch_kernel(minmax, execution_policy(nparts, raw::DYNAMIC_QUANTIZE_BLOCK, stream),
+                      input.data().get(), input.size(), pmin, pmax);
+        auto params = raw::dynamic_quantize_params;
+        launch_kernel(params, execution_policy(1, raw::DYNAMIC_QUANTIZE_BLOCK, stream),
+                      pmin, pmax, static_cast<int>(nparts), scale.data().get(), zero_point.data().get());
+    }
+
+    template void dequantize_linear<std::int8_t>(const Stream&, Span<float>, View<std::int8_t>, View<float>, View<std::int8_t>, std::size_t, std::size_t);
+    template void dequantize_linear<std::uint8_t>(const Stream&, Span<float>, View<std::uint8_t>, View<float>, View<std::uint8_t>, std::size_t, std::size_t);
+    template void quantize_linear<std::int8_t>(const Stream&, Span<std::int8_t>, View<float>, View<float>, View<std::int8_t>, std::size_t, std::size_t);
+    template void quantize_linear<std::uint8_t>(const Stream&, Span<std::uint8_t>, View<float>, View<float>, View<std::uint8_t>, std::size_t, std::size_t);
+
+}}}} /* namespace cv::dnn::cuda4dnn::kernels */

modules/dnn/src/cuda/range.cu

@@ -0,0 +1,72 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+// Copyright (C) 2026, BigVision LLC, all rights reserved.
+// Third party copyrights are property of their respective owners.
+#include <cuda_runtime.h>
+#include <cuda_fp16.h>
+
+#include "types.hpp"
+#include "grid_stride_range.hpp"
+#include "execution.hpp"
+
+#include "../cuda4dnn/csl/stream.hpp"
+#include "../cuda4dnn/csl/span.hpp"
+
+#include "../cuda4dnn/kernels/range.hpp"
+
+#include <cstdint>
+
+using namespace cv::dnn::cuda4dnn::csl;
+using namespace cv::dnn::cuda4dnn::csl::device;
+
+namespace cv { namespace dnn { namespace cuda4dnn { namespace kernels {
+
+    namespace raw {
+        __device__ inline double as_double(float v)         { return static_cast<double>(v); }
+        __device__ inline double as_double(__half v)        { return static_cast<double>(__half2float(v)); }
+        __device__ inline double as_double(std::int8_t v)   { return static_cast<double>(v); }
+        __device__ inline double as_double(std::uint8_t v)  { return static_cast<double>(v); }
+        __device__ inline double as_double(std::int32_t v)  { return static_cast<double>(v); }
+        __device__ inline double as_double(std::int64_t v)  { return static_cast<double>(v); }
+
+        template <class T>
+        __device__ inline T range_value(T start, T delta, std::size_t i) {
+            return static_cast<T>(as_double(start) + static_cast<double>(i) * as_double(delta));
+        }
+
+        // int64 stays in integer arithmetic: double holds integers exactly only up to 2^53
+        __device__ inline std::int64_t range_value(std::int64_t start, std::int64_t delta, std::size_t i) {
+            return start + static_cast<std::int64_t>(i) * delta;
+        }
+
+        template <class T>
+        __global__ void range(Span<T> output, const T* start_p, const T* delta_p)
+        {
+            const T start = *start_p;
+            const T delta = *delta_p;
+            for (auto i : grid_stride_range(output.size()))
+                output[i] = range_value(start, delta, static_cast<std::size_t>(i));
+        }
+    }
+
+    template <class T>
+    void range(const Stream& stream, Span<T> output, const T* start, const T* delta)
+    {
+        if (output.size() == 0)
+            return;
+        auto kernel = raw::range<T>;
+        auto policy = make_policy(kernel, output.size(), 0, stream);
+        launch_kernel(kernel, policy, output, start, delta);
+    }
+
+#if !defined(__CUDA_ARCH__) || (__CUDA_ARCH__ >= 530)
+    template void range(const Stream&, Span<__half>, const __half*, const __half*);
+#endif
+    template void range(const Stream&, Span<float>, const float*, const float*);
+    template void range(const Stream&, Span<int8_t>, const int8_t*, const int8_t*);
+    template void range(const Stream&, Span<uint8_t>, const uint8_t*, const uint8_t*);
+    template void range(const Stream&, Span<int32_t>, const int32_t*, const int32_t*);
+    template void range(const Stream&, Span<int64_t>, const int64_t*, const int64_t*);
+
+}}}} /* namespace cv::dnn::cuda4dnn::kernels */

modules/dnn/src/cuda/resize.cu

@@ -242,4 +242,76 @@ namespace cv { namespace dnn { namespace cuda4dnn { namespace kernels {
 #endif
     template void resize_bilinear<float>(const Stream&, TensorSpan<float>, TensorView<float>, float, float, bool);
 
+    namespace raw {
+        template <class T>
+        __global__ void resize_separable(Span<T> output, View<T> input, size_type in_h, size_type in_w,
+                                         size_type out_h, size_type out_w,
+                                         View<int> idx_y, View<float> w_y, View<int> oob_y, int taps_y,
+                                         View<int> idx_x, View<float> w_x, View<int> oob_x, int taps_x,
+                                         float extrapolation_value)
+        {
+            for (auto i : grid_stride_range(output.size())) {
+                const index_type ox = i % out_w;
+                const index_type oy = (i / out_w) % out_h;
+                const index_type plane = i / (out_w * out_h);
+                if (oob_y[oy] || oob_x[ox]) {
+                    output[i] = static_cast<T>(extrapolation_value);
+                    continue;
+                }
+                const index_type in_plane = plane * in_h * in_w;
+                float acc = 0.f;
+                for (int ky = 0; ky < taps_y; ky++) {
+                    const int iy = idx_y[oy * taps_y + ky];
+                    if (iy < 0)
+                        continue;
+                    float row = 0.f;
+                    for (int kx = 0; kx < taps_x; kx++) {
+                        const int ix = idx_x[ox * taps_x + kx];
+                        if (ix < 0)
+                            continue;
+                        row += w_x[ox * taps_x + kx] * static_cast<float>(input[in_plane + iy * in_w + ix]);
+                    }
+                    acc += w_y[oy * taps_y + ky] * row;
+                }
+                output[i] = static_cast<T>(acc);
+            }
+        }
+    }
+
+    template <class T>
+    void resize_separable(const Stream& stream, TensorSpan<T> output, TensorView<T> input,
+                          View<int> idx_y, View<float> w_y, View<int> oob_y, int taps_y,
+                          View<int> idx_x, View<float> w_x, View<int> oob_x, int taps_x,
+                          float extrapolation_value)
+    {
+        CV_Assert(input.rank() == output.rank() && input.rank() >= 2);
+        const auto out_h = output.get_axis_size(-2), out_w = output.get_axis_size(-1);
+        const auto in_h = input.get_axis_size(-2), in_w = input.get_axis_size(-1);
+        CV_Assert(input.size() / (in_h * in_w) == output.size() / (out_h * out_w));
+        CV_Assert(idx_y.size() == out_h * taps_y && idx_x.size() == out_w * taps_x);
+        if (output.size() == 0)
+            return;
+
+        auto kernel = raw::resize_separable<T>;
+        auto policy = make_policy(kernel, output.size(), 0, stream);
+        launch_kernel(kernel, policy, Span<T>(output.get(), output.size()), View<T>(input.get(), input.size()),
+                      in_h, in_w, out_h, out_w, idx_y, w_y, oob_y, taps_y, idx_x, w_x, oob_x, taps_x,
+                      extrapolation_value);
+    }
+
+#if !defined(__CUDA_ARCH__) || (__CUDA_ARCH__ >= 530)
+    template void resize_separable<__half>(const Stream&, TensorSpan<__half>, TensorView<__half>,
+        View<int>, View<float>, View<int>, int, View<int>, View<float>, View<int>, int, float);
+#endif
+    template void resize_separable<float>(const Stream&, TensorSpan<float>, TensorView<float>,
+        View<int>, View<float>, View<int>, int, View<int>, View<float>, View<int>, int, float);
+    template void resize_separable<int8_t>(const Stream&, TensorSpan<int8_t>, TensorView<int8_t>,
+        View<int>, View<float>, View<int>, int, View<int>, View<float>, View<int>, int, float);
+    template void resize_separable<uint8_t>(const Stream&, TensorSpan<uint8_t>, TensorView<uint8_t>,
+        View<int>, View<float>, View<int>, int, View<int>, View<float>, View<int>, int, float);
+    template void resize_separable<int32_t>(const Stream&, TensorSpan<int32_t>, TensorView<int32_t>,
+        View<int>, View<float>, View<int>, int, View<int>, View<float>, View<int>, int, float);
+    template void resize_separable<int64_t>(const Stream&, TensorSpan<int64_t>, TensorView<int64_t>,
+        View<int>, View<float>, View<int>, int, View<int>, View<float>, View<int>, int, float);
+
 }}}} /* namespace cv::dnn::cuda4dnn::kernels */

modules/dnn/src/cuda/rms_norm.cu

@@ -0,0 +1,97 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+// Copyright (C) 2026, BigVision LLC, all rights reserved.
+// Third party copyrights are property of their respective owners.
+#include <cuda_runtime.h>
+#include <cuda_fp16.h>
+
+#include <cub/block/block_reduce.cuh>
+
+#include "types.hpp"
+#include "execution.hpp"
+
+#include "../cuda4dnn/csl/stream.hpp"
+#include "../cuda4dnn/csl/span.hpp"
+
+#include "../cuda4dnn/kernels/rms_norm.hpp"
+
+#include <opencv2/core.hpp>
+
+#include <algorithm>
+#include <cstddef>
+#include <cstdint>
+
+using namespace cv::dnn::cuda4dnn::csl;
+using namespace cv::dnn::cuda4dnn::csl::device;
+
+namespace cv { namespace dnn { namespace cuda4dnn { namespace kernels {
+
+    namespace raw {
+
+        template <class T, int BlockSize>
+        __global__ void rms_norm(Span<T> output, View<T> input, View<T> scale,
+                                 std::size_t loops, std::size_t norm_size, float epsilon)
+        {
+            using BlockReduce = cub::BlockReduce<float, BlockSize>;
+            __shared__ typename BlockReduce::TempStorage reduce_storage;
+            __shared__ float inv_stddev;
+
+            const T* in = input.data().get();
+            const T* sc = scale.data().get();
+            T* out = output.data().get();
+
+            for (std::size_t row = blockIdx.x; row < loops; row += gridDim.x) {
+                const std::int64_t base = static_cast<std::int64_t>(row) * norm_size;
+
+                float thread_sum_sq = 0.f;
+                for (std::size_t j = threadIdx.x; j < norm_size; j += BlockSize) {
+                    const float v = static_cast<float>(in[base + j]);
+                    thread_sum_sq += v * v;
+                }
+
+                const float block_sum_sq = BlockReduce(reduce_storage).Sum(thread_sum_sq);
+
+                if (threadIdx.x == 0) {
+                    // recenter=false collapses the variance to the mean square
+                    const float mean_square = block_sum_sq / static_cast<float>(norm_size);
+                    inv_stddev = 1.f / sqrtf(fmaxf(0.f, mean_square) + epsilon);
+                }
+                __syncthreads();  // publishes inv_stddev and frees reduce_storage for the next row
+
+                for (std::size_t j = threadIdx.x; j < norm_size; j += BlockSize) {
+                    const float v = static_cast<float>(in[base + j]);
+                    const float s = static_cast<float>(sc[j]);
+                    out[base + j] = static_cast<T>(s * v * inv_stddev);
+                }
+                __syncthreads();
+            }
+        }
+    }
+
+    template <class T>
+    void rms_norm(const Stream& stream,
+        Span<T> output, View<T> input, View<T> scale,
+        std::size_t loops, std::size_t norm_size, float epsilon)
+    {
+        CV_Assert(loops > 0 && norm_size > 0);
+        CV_Assert(scale.size() >= norm_size);
+
+        constexpr int block_size = 256;
+        constexpr int max_blocks = 65535;
+        const int grid_size = static_cast<int>(std::min<std::size_t>(max_blocks, loops));
+
+        raw::rms_norm<T, block_size><<<grid_size, block_size, 0, stream.get()>>>(
+            output, input, scale, loops, norm_size, epsilon);
+    }
+
+#if !defined(__CUDA_ARCH__) || (__CUDA_ARCH__ >= 530)
+    template void rms_norm(const Stream&, Span<__half>, View<__half>, View<__half>, std::size_t, std::size_t, float);
+#endif
+    template void rms_norm(const Stream&, Span<float>, View<float>, View<float>, std::size_t, std::size_t, float);
+    template void rms_norm(const Stream&, Span<int8_t>, View<int8_t>, View<int8_t>, std::size_t, std::size_t, float);
+    template void rms_norm(const Stream&, Span<uint8_t>, View<uint8_t>, View<uint8_t>, std::size_t, std::size_t, float);
+    template void rms_norm(const Stream&, Span<int32_t>, View<int32_t>, View<int32_t>, std::size_t, std::size_t, float);
+    template void rms_norm(const Stream&, Span<int64_t>, View<int64_t>, View<int64_t>, std::size_t, std::size_t, float);
+
+}}}} /* namespace cv::dnn::cuda4dnn::kernels */

modules/dnn/src/cuda/roi_align.cu

@@ -0,0 +1,161 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+// Copyright (C) 2026, BigVision LLC, all rights reserved.
+// Third party copyrights are property of their respective owners.
+#include <cuda_runtime.h>
+#include <cuda_fp16.h>
+
+#include "types.hpp"
+#include "grid_stride_range.hpp"
+#include "math.hpp"
+#include "execution.hpp"
+
+#include "../cuda4dnn/csl/stream.hpp"
+#include "../cuda4dnn/csl/span.hpp"
+
+#include "../cuda4dnn/kernels/roi_align.hpp"
+
+#include <opencv2/core.hpp>
+
+#include <cfloat>
+#include <cstddef>
+#include <cstdint>
+
+using namespace cv::dnn::cuda4dnn::csl;
+using namespace cv::dnn::cuda4dnn::csl::device;
+
+namespace cv { namespace dnn { namespace cuda4dnn { namespace kernels {
+
+    namespace raw {
+
+        __device__ inline std::int64_t read_batch_index(const void* data, std::size_t element_size, int offset) {
+            if (element_size == sizeof(std::int32_t))
+                return *(reinterpret_cast<const std::int32_t*>(data) + offset);
+            return *(reinterpret_cast<const std::int64_t*>(data) + offset);
+        }
+
+        /* One thread owns one output element. The ROI geometry is recomputed per thread
+         * rather than cached per ROI (as the CPU path does); it is a handful of flops
+         * against a memory-bound gather, and it keeps the kernel free of shared state. */
+        template <class T>
+        __global__ void roi_align(Span<T> output, View<T> input, View<float> rois,
+                                  const void* batch_indices, std::size_t batch_index_element_size,
+                                  RoiAlignParams p)
+        {
+            for (auto id : grid_stride_range(output.size())) {
+                int rem = id;
+                const int pw = rem % p.output_width;  rem /= p.output_width;
+                const int ph = rem % p.output_height; rem /= p.output_height;
+                const int c  = rem % p.channels;      rem /= p.channels;
+                const int r  = rem;
+
+                const float* roi = rois.data().get() + static_cast<std::int64_t>(r) * 4;
+                const float x1 = roi[0] * p.spatial_scale - p.offset;
+                const float y1 = roi[1] * p.spatial_scale - p.offset;
+                const float x2 = roi[2] * p.spatial_scale - p.offset;
+                const float y2 = roi[3] * p.spatial_scale - p.offset;
+
+                float roi_width  = x2 - x1;
+                float roi_height = y2 - y1;
+                if (p.clamp_malformed_roi) {
+                    roi_width  = fmaxf(roi_width, 1.f);
+                    roi_height = fmaxf(roi_height, 1.f);
+                }
+
+                const float bin_size_w = roi_width / p.output_width;
+                const float bin_size_h = roi_height / p.output_height;
+
+                const int gw = device::max<int>((p.sampling_ratio > 0) ? p.sampling_ratio
+                                                                       : static_cast<int>(ceilf(roi_width / p.output_width)), 1);
+                const int gh = device::max<int>((p.sampling_ratio > 0) ? p.sampling_ratio
+                                                                       : static_cast<int>(ceilf(roi_height / p.output_height)), 1);
+                const int sample_count = gw * gh;
+                const float step_y = bin_size_h / gh;
+                const float step_x = bin_size_w / gw;
+
+                const std::int64_t b = read_batch_index(batch_indices, batch_index_element_size, r);
+                // a kernel can't raise the CPU path's error for an index outside the batch,
+                // so it writes 0 instead of reading outside the input
+                if (b < 0 || b >= p.batch_size) {
+                    output[id] = static_cast<T>(0.f);
+                    continue;
+                }
+                const T* img = input.data().get() +
+                               ((b * p.channels) + c) * static_cast<std::int64_t>(p.height) * p.width;
+
+                float outv = p.max_mode ? -FLT_MAX : 0.f;
+                for (int s = 0; s < sample_count; ++s) {
+                    const int iy = s / gw;
+                    const int ix = s - iy * gw;
+                    const float yy = y1 + ph * bin_size_h + (iy + 0.5f) * step_y;
+                    const float xx = x1 + pw * bin_size_w + (ix + 0.5f) * step_x;
+
+                    /* precomputeBilinearSample() returns an all-zero sample when the point
+                     * is fully outside [-1, extent], so it contributes 0 in both modes. */
+                    float v = 0.f;
+                    if (!(yy < -1.f || yy > p.height || xx < -1.f || xx > p.width)) {
+                        const float cy = fminf(fmaxf(yy, 0.f), static_cast<float>(p.height - 1));
+                        const float cx = fminf(fmaxf(xx, 0.f), static_cast<float>(p.width - 1));
+
+                        const int y_low = static_cast<int>(cy);
+                        const int x_low = static_cast<int>(cx);
+                        const int y_high = device::min<int>(y_low + 1, p.height - 1);
+                        const int x_high = device::min<int>(x_low + 1, p.width - 1);
+
+                        const float ly = cy - y_low, lx = cx - x_low;
+                        const float hy = 1.f - ly,   hx = 1.f - lx;
+
+                        const std::int64_t row_low  = static_cast<std::int64_t>(y_low) * p.width;
+                        const std::int64_t row_high = static_cast<std::int64_t>(y_high) * p.width;
+
+                        const float v00 = static_cast<float>(img[row_low  + x_low]);
+                        const float v01 = static_cast<float>(img[row_low  + x_high]);
+                        const float v10 = static_cast<float>(img[row_high + x_low]);
+                        const float v11 = static_cast<float>(img[row_high + x_high]);
+
+                        v = p.max_mode
+                                ? fmaxf(fmaxf(hy * hx * v00, hy * lx * v01),
+                                        fmaxf(ly * hx * v10, ly * lx * v11))
+                                : (hy * hx * v00 + hy * lx * v01 + ly * hx * v10 + ly * lx * v11);
+                    }
+
+                    outv = p.max_mode ? fmaxf(outv, v) : (outv + v);
+                }
+
+                outv = p.max_mode ? ((outv == -FLT_MAX) ? 0.f : outv)
+                                  : (outv / static_cast<float>(sample_count));
+                output[id] = static_cast<T>(outv);
+            }
+        }
+    }
+
+    template <class T>
+    void roi_align(const Stream& stream,
+        Span<T> output, View<T> input, View<float> rois,
+        const void* batch_indices, std::size_t batch_index_element_size,
+        const RoiAlignParams& params)
+    {
+        CV_Assert(batch_index_element_size == sizeof(std::int32_t) ||
+                  batch_index_element_size == sizeof(std::int64_t));
+        CV_Assert(params.output_height > 0 && params.output_width > 0);
+        CV_Assert(params.batch_size > 0 && params.height > 0 && params.width > 0 && params.channels > 0);
+
+        if (output.size() == 0)
+            return;
+
+        auto kernel = raw::roi_align<T>;
+        auto policy = make_policy(kernel, output.size(), 0, stream);
+        launch_kernel(kernel, policy, output, input, rois, batch_indices, batch_index_element_size, params);
+    }
+
+#if !defined(__CUDA_ARCH__) || (__CUDA_ARCH__ >= 530)
+    template void roi_align(const Stream&, Span<__half>, View<__half>, View<float>, const void*, std::size_t, const RoiAlignParams&);
+#endif
+    template void roi_align(const Stream&, Span<float>, View<float>, View<float>, const void*, std::size_t, const RoiAlignParams&);
+    template void roi_align(const Stream&, Span<int8_t>, View<int8_t>, View<float>, const void*, std::size_t, const RoiAlignParams&);
+    template void roi_align(const Stream&, Span<uint8_t>, View<uint8_t>, View<float>, const void*, std::size_t, const RoiAlignParams&);
+    template void roi_align(const Stream&, Span<int32_t>, View<int32_t>, View<float>, const void*, std::size_t, const RoiAlignParams&);
+    template void roi_align(const Stream&, Span<int64_t>, View<int64_t>, View<float>, const void*, std::size_t, const RoiAlignParams&);
+
+}}}} /* namespace cv::dnn::cuda4dnn::kernels */

modules/dnn/src/cuda/scatter_elements.cu

@@ -0,0 +1,107 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+// Copyright (C) 2026, BigVision LLC, all rights reserved.
+// Third party copyrights are property of their respective owners.
+#include <cuda_runtime.h>
+#include <cuda_fp16.h>
+
+#include "array.hpp"
+#include "types.hpp"
+#include "grid_stride_range.hpp"
+#include "execution.hpp"
+
+#include "../cuda4dnn/csl/stream.hpp"
+#include "../cuda4dnn/csl/span.hpp"
+#include "../cuda4dnn/csl/tensor.hpp"
+
+#include "../cuda4dnn/kernels/scatter_elements.hpp"
+
+#include <opencv2/core.hpp>
+
+#include <cstdint>
+#include <vector>
+
+using namespace cv::dnn::cuda4dnn::csl;
+using namespace cv::dnn::cuda4dnn::csl::device;
+
+namespace cv { namespace dnn { namespace cuda4dnn { namespace kernels {
+
+    namespace raw {
+        template <class T, class TIdx>
+        __global__ void scatter_elements(
+            Span<T> output, View<TIdx> indices, View<T> updates,
+            int rank, int axis,
+            array<std::int64_t, CSL_MAX_TENSOR_RANK> indices_dims,
+            array<std::int64_t, CSL_MAX_TENSOR_RANK> output_strides,
+            std::int64_t axis_dim_size)
+        {
+            for (auto id : grid_stride_range(updates.size())) {
+                std::int64_t rem = static_cast<std::int64_t>(id);
+                std::int64_t offset = 0;
+
+                for (int d = rank - 1; d >= 0; --d) {
+                    const std::int64_t dim = indices_dims[d];
+                    const std::int64_t coord = rem % dim;
+                    rem /= dim;
+
+                    std::int64_t use = coord;
+                    if (d == axis) {
+                        std::int64_t idx = static_cast<std::int64_t>(indices.data().get()[id]);
+                        if (idx < 0)
+                            idx += axis_dim_size;
+                        if (idx < 0)
+                            idx = 0;
+                        if (idx >= axis_dim_size)
+                            idx = axis_dim_size - 1;
+                        use = idx;
+                    }
+                    offset += use * output_strides[d];
+                }
+
+                output[offset] = updates.data().get()[id];
+            }
+        }
+    }
+
+    template <class T, class TIdx>
+    void scatter_elements(const Stream& stream,
+        Span<T> output, View<TIdx> indices, View<T> updates,
+        int rank, int axis,
+        const std::vector<std::int64_t>& indices_dims,
+        const std::vector<std::int64_t>& output_strides,
+        std::int64_t axis_dim_size)
+    {
+        CV_Assert(rank <= CSL_MAX_TENSOR_RANK);
+        CV_Assert(static_cast<int>(indices_dims.size()) == rank);
+        CV_Assert(static_cast<int>(output_strides.size()) == rank);
+
+        if (updates.size() == 0)
+            return;
+
+        array<std::int64_t, CSL_MAX_TENSOR_RANK> dims_k, strides_k;
+        dims_k.assign(std::begin(indices_dims), std::end(indices_dims));
+        strides_k.assign(std::begin(output_strides), std::end(output_strides));
+
+        auto kernel = raw::scatter_elements<T, TIdx>;
+        auto policy = make_policy(kernel, updates.size(), 0, stream);
+        launch_kernel(kernel, policy, output, indices, updates, rank, axis,
+                      dims_k, strides_k, axis_dim_size);
+    }
+
+#define CV_SCATTER_ELEM_INST(T) \
+    template void scatter_elements(const Stream&, Span<T>, View<std::int32_t>, View<T>, int, int, const std::vector<std::int64_t>&, const std::vector<std::int64_t>&, std::int64_t); \
+    template void scatter_elements(const Stream&, Span<T>, View<std::int64_t>, View<T>, int, int, const std::vector<std::int64_t>&, const std::vector<std::int64_t>&, std::int64_t);
+
+#if !defined(__CUDA_ARCH__) || (__CUDA_ARCH__ >= 530)
+    CV_SCATTER_ELEM_INST(__half)
+#endif
+    CV_SCATTER_ELEM_INST(float)
+    CV_SCATTER_ELEM_INST(int8_t)
+    CV_SCATTER_ELEM_INST(uint8_t)
+    CV_SCATTER_ELEM_INST(int32_t)
+    CV_SCATTER_ELEM_INST(int64_t)
+
+#undef CV_SCATTER_ELEM_INST
+
+}}}} /* namespace cv::dnn::cuda4dnn::kernels */

modules/dnn/src/cuda/scatter_nd.cu

@@ -0,0 +1,117 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+// Copyright (C) 2026, BigVision LLC, all rights reserved.
+// Third party copyrights are property of their respective owners.
+#include <cuda_runtime.h>
+#include <cuda_fp16.h>
+
+#include "array.hpp"
+#include "types.hpp"
+#include "grid_stride_range.hpp"
+#include "execution.hpp"
+
+#include "../cuda4dnn/csl/stream.hpp"
+#include "../cuda4dnn/csl/span.hpp"
+#include "../cuda4dnn/csl/tensor.hpp"
+
+#include "../cuda4dnn/kernels/scatter_nd.hpp"
+
+#include <opencv2/core.hpp>
+
+#include <cstddef>
+#include <cstdint>
+#include <vector>
+
+using namespace cv::dnn::cuda4dnn::csl;
+using namespace cv::dnn::cuda4dnn::csl::device;
+
+namespace cv { namespace dnn { namespace cuda4dnn { namespace kernels {
+
+    namespace raw {
+        template <class T, class TIdx>
+        __global__ void scatter_nd(
+            Span<T> output, View<TIdx> indices, View<T> updates,
+            size_type num_indices, std::int64_t last_index_dimension,
+            array<std::int64_t, CSL_MAX_TENSOR_RANK> element_counts,
+            array<std::int64_t, CSL_MAX_TENSOR_RANK> input_dims,
+            std::int64_t num_updates_elements)
+        {
+            for (auto id : grid_stride_range(num_indices)) {
+                std::int64_t data_offset = 0;
+
+                std::int64_t indices_start = last_index_dimension * id;
+                for (std::int64_t j = 0; j < last_index_dimension; ++j) {
+                    std::int64_t index = static_cast<std::int64_t>(indices.data().get()[indices_start + j]);
+
+                    std::int64_t element_count_dim = element_counts[j];
+                    std::int64_t dim_value = input_dims[j];
+
+                    if (index >= 0) {
+                        if (index >= dim_value) {
+                            index = dim_value - 1;
+                        }
+                    } else {
+                        if (index < -dim_value) {
+                            index = 0;
+                        } else {
+                            index += dim_value;
+                        }
+                    }
+
+                    data_offset += (index * element_count_dim);
+                }
+
+                const T* updates_data_base = updates.data().get() + num_updates_elements * static_cast<std::int64_t>(id);
+                T* output_data_base = output.data().get() + data_offset;
+
+                for (std::int64_t i = 0; i < num_updates_elements; ++i) {
+                    output_data_base[i] = updates_data_base[i];
+                }
+            }
+        }
+    }
+
+    template <class T, class TIdx>
+    void scatter_nd(const Stream& stream,
+        Span<T> output, View<TIdx> indices, View<T> updates,
+        std::size_t num_indices, std::int64_t last_index_dimension,
+        const std::vector<std::int64_t>& element_counts,
+        const std::vector<std::int64_t>& input_dims,
+        std::size_t num_updates_elements)
+    {
+        CV_Assert(last_index_dimension <= CSL_MAX_TENSOR_RANK);
+        CV_Assert(element_counts.size() == static_cast<std::size_t>(last_index_dimension));
+        CV_Assert(input_dims.size() == static_cast<std::size_t>(last_index_dimension));
+
+        if (num_indices == 0)
+            return;
+
+        array<std::int64_t, CSL_MAX_TENSOR_RANK> element_counts_k, input_dims_k;
+        element_counts_k.assign(std::begin(element_counts), std::end(element_counts));
+        input_dims_k.assign(std::begin(input_dims), std::end(input_dims));
+
+        auto kernel = raw::scatter_nd<T, TIdx>;
+        auto policy = make_policy(kernel, num_indices, 0, stream);
+        launch_kernel(kernel, policy, output, indices, updates,
+                      static_cast<size_type>(num_indices), last_index_dimension,
+                      element_counts_k, input_dims_k,
+                      static_cast<std::int64_t>(num_updates_elements));
+    }
+
+#if !defined(__CUDA_ARCH__) || (__CUDA_ARCH__ >= 530)
+    template void scatter_nd(const Stream&, Span<__half>, View<std::int32_t>, View<__half>, std::size_t, std::int64_t, const std::vector<std::int64_t>&, const std::vector<std::int64_t>&, std::size_t);
+    template void scatter_nd(const Stream&, Span<__half>, View<std::int64_t>, View<__half>, std::size_t, std::int64_t, const std::vector<std::int64_t>&, const std::vector<std::int64_t>&, std::size_t);
+#endif
+    template void scatter_nd(const Stream&, Span<float>, View<std::int32_t>, View<float>, std::size_t, std::int64_t, const std::vector<std::int64_t>&, const std::vector<std::int64_t>&, std::size_t);
+    template void scatter_nd(const Stream&, Span<float>, View<std::int64_t>, View<float>, std::size_t, std::int64_t, const std::vector<std::int64_t>&, const std::vector<std::int64_t>&, std::size_t);
+    template void scatter_nd(const Stream&, Span<int8_t>, View<std::int32_t>, View<int8_t>, std::size_t, std::int64_t, const std::vector<std::int64_t>&, const std::vector<std::int64_t>&, std::size_t);
+    template void scatter_nd(const Stream&, Span<int8_t>, View<std::int64_t>, View<int8_t>, std::size_t, std::int64_t, const std::vector<std::int64_t>&, const std::vector<std::int64_t>&, std::size_t);
+    template void scatter_nd(const Stream&, Span<uint8_t>, View<std::int32_t>, View<uint8_t>, std::size_t, std::int64_t, const std::vector<std::int64_t>&, const std::vector<std::int64_t>&, std::size_t);
+    template void scatter_nd(const Stream&, Span<uint8_t>, View<std::int64_t>, View<uint8_t>, std::size_t, std::int64_t, const std::vector<std::int64_t>&, const std::vector<std::int64_t>&, std::size_t);
+    template void scatter_nd(const Stream&, Span<int32_t>, View<std::int32_t>, View<int32_t>, std::size_t, std::int64_t, const std::vector<std::int64_t>&, const std::vector<std::int64_t>&, std::size_t);
+    template void scatter_nd(const Stream&, Span<int32_t>, View<std::int64_t>, View<int32_t>, std::size_t, std::int64_t, const std::vector<std::int64_t>&, const std::vector<std::int64_t>&, std::size_t);
+    template void scatter_nd(const Stream&, Span<int64_t>, View<std::int32_t>, View<int64_t>, std::size_t, std::int64_t, const std::vector<std::int64_t>&, const std::vector<std::int64_t>&, std::size_t);
+    template void scatter_nd(const Stream&, Span<int64_t>, View<std::int64_t>, View<int64_t>, std::size_t, std::int64_t, const std::vector<std::int64_t>&, const std::vector<std::int64_t>&, std::size_t);
+
+}}}} /* namespace cv::dnn::cuda4dnn::kernels */

modules/dnn/src/cuda/topk.cu

@@ -0,0 +1,287 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+// Copyright (C) 2026, BigVision LLC, all rights reserved.
+// Third party copyrights are property of their respective owners.
+#include <cuda_runtime.h>
+#include <cuda_fp16.h>
+
+#include <cub/block/block_reduce.cuh>
+#include <cub/device/device_radix_sort.cuh>
+#include <cub/device/device_segmented_sort.cuh>
+
+#include "types.hpp"
+#include "execution.hpp"
+
+#include "../cuda4dnn/csl/stream.hpp"
+#include "../cuda4dnn/csl/span.hpp"
+
+#include "../cuda4dnn/kernels/topk.hpp"
+
+#include <opencv2/core.hpp>
+
+#include <cuda/std/limits>
+
+#include <algorithm>
+#include <cstdint>
+
+using namespace cv::dnn::cuda4dnn::csl;
+using namespace cv::dnn::cuda4dnn::csl::device;
+
+namespace cv { namespace dnn { namespace cuda4dnn { namespace kernels {
+
+    namespace raw {
+
+        /* float is exact for __half, float, int8 and uint8; int32/int64 compare in
+         * their own type, since float holds integers exactly only up to 2^24. */
+        template <class T> struct TopKCompare { using type = float; };
+        template <> struct TopKCompare<std::int32_t> { using type = std::int32_t; };
+        template <> struct TopKCompare<std::int64_t> { using type = std::int64_t; };
+
+        /* Starting value that any real element beats: infinity for float types,
+         * and the lowest or highest integer for int32/int64. */
+        template <class C>
+        __device__ inline C topk_worst(bool largest) {
+            using limits = ::cuda::std::numeric_limits<C>;
+            if (limits::has_infinity)
+                return largest ? -limits::infinity() : limits::infinity();
+            return largest ? limits::lowest() : limits::max();
+        }
+
+        template <class C>
+        struct Candidate {
+            C value;
+            int index;
+        };
+
+        /* Strict weak ordering matching topk2_layer.cpp's comparators: better value
+         * first, and on an exact tie the smaller index. */
+        template <class C>
+        __device__ inline bool better_than(const Candidate<C>& a, const Candidate<C>& b, bool largest) {
+            if (a.value == b.value)
+                return a.index < b.index;
+            return largest ? (a.value > b.value) : (a.value < b.value);
+        }
+
+        template <class C>
+        struct MergeCandidate {
+            bool largest;
+            __device__ inline Candidate<C> operator()(const Candidate<C>& a, const Candidate<C>& b) const {
+                return better_than(a, b, largest) ? a : b;
+            }
+        };
+
+        template <class T, int BlockSize>
+        __global__ void topk(Span<T> values, Span<std::int64_t> indices, View<T> input,
+                             int outer, int dim_axis, int inner, int k, bool largest)
+        {
+            using C = typename TopKCompare<T>::type;
+            using Cand = Candidate<C>;
+            using BlockReduce = cub::BlockReduce<Cand, BlockSize>;
+            __shared__ typename BlockReduce::TempStorage reduce_storage;
+            __shared__ Cand selected;
+
+            const int nslices = outer * inner;
+            const C worst = topk_worst<C>(largest);
+
+            for (int slice = blockIdx.x; slice < nslices; slice += gridDim.x) {
+                const int b = slice / inner;
+                const int j = slice - b * inner;
+
+                const std::int64_t in_base  = static_cast<std::int64_t>(b) * dim_axis * inner + j;
+                const std::int64_t out_base = static_cast<std::int64_t>(b) * k * inner + j;
+
+                for (int rank = 0; rank < k; ++rank) {
+                    Cand best{worst, ::cuda::std::numeric_limits<int>::max()};
+
+                    for (int u = threadIdx.x; u < dim_axis; u += BlockSize) {
+                        const C v = static_cast<C>(input.data().get()[in_base + static_cast<std::int64_t>(u) * inner]);
+                        const Cand cand{v, u};
+
+                        /* Everything at or before the previous pick is spent. Ordering
+                         * strictly means no element can be chosen twice, and duplicates
+                         * are still all reachable because index breaks the tie. */
+                        if (rank > 0 && !better_than(selected, cand, largest))
+                            continue;
+
+                        if (better_than(cand, best, largest))
+                            best = cand;
+                    }
+
+                    const Cand winner = BlockReduce(reduce_storage).Reduce(best, MergeCandidate<C>{largest});
+
+                    if (threadIdx.x == 0) {
+                        selected = winner;
+                        const std::int64_t out = out_base + static_cast<std::int64_t>(rank) * inner;
+                        values[out] = static_cast<T>(winner.value);
+                        indices[out] = static_cast<std::int64_t>(winner.index);
+                    }
+                    __syncthreads();  // publishes `selected` and frees reduce_storage for the next rank
+                }
+            }
+        }
+
+        __global__ void topk_sort_init(int* ids, int* offsets, int n, int nsegments, int dim_axis)
+        {
+            const int stride = gridDim.x * blockDim.x;
+            for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < n; i += stride)
+                ids[i] = i % dim_axis;
+            for (int i = blockIdx.x * blockDim.x + threadIdx.x; i <= nsegments; i += stride)
+                offsets[i] = i * dim_axis;
+        }
+
+        template <class T>
+        __global__ void topk_sort_gather(Span<T> values, Span<std::int64_t> indices,
+                                         const T* keys, const int* ids, int dim_axis, int k)
+        {
+            const int m = static_cast<int>(values.size());
+            const int stride = gridDim.x * blockDim.x;
+            for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < m; i += stride) {
+                const int b = i / k;
+                const int src = b * dim_axis + (i - b * k);
+                values[i] = keys[src];
+                indices[i] = static_cast<std::int64_t>(ids[src]);
+            }
+        }
+    }
+
+    namespace detail {
+        struct TopKSortLayout {
+            std::size_t keys, ids_in, ids_out, offsets, temp, total;
+        };
+
+        inline std::size_t alignTo256(std::size_t bytes)
+        {
+            return (bytes + 255) & ~static_cast<std::size_t>(255);
+        }
+
+        template <class T>
+        TopKSortLayout topkSortLayout(int outer, int dim_axis, bool largest, cudaStream_t stream)
+        {
+            const std::int64_t n = static_cast<std::int64_t>(outer) * dim_axis;
+            std::size_t temp_bytes = 0;
+            if (outer == 1 && largest)
+                CUDA4DNN_CHECK_CUDA(cub::DeviceRadixSort::SortPairsDescending(
+                    nullptr, temp_bytes, static_cast<const T*>(nullptr), static_cast<T*>(nullptr),
+                    static_cast<const int*>(nullptr), static_cast<int*>(nullptr), n, 0,
+                    static_cast<int>(sizeof(T) * 8), stream));
+            else if (outer == 1)
+                CUDA4DNN_CHECK_CUDA(cub::DeviceRadixSort::SortPairs(
+                    nullptr, temp_bytes, static_cast<const T*>(nullptr), static_cast<T*>(nullptr),
+                    static_cast<const int*>(nullptr), static_cast<int*>(nullptr), n, 0,
+                    static_cast<int>(sizeof(T) * 8), stream));
+            else if (largest)
+                CUDA4DNN_CHECK_CUDA(cub::DeviceSegmentedSort::StableSortPairsDescending(
+                    nullptr, temp_bytes, static_cast<const T*>(nullptr), static_cast<T*>(nullptr),
+                    static_cast<const int*>(nullptr), static_cast<int*>(nullptr), n, outer,
+                    static_cast<const int*>(nullptr), static_cast<const int*>(nullptr), stream));
+            else
+                CUDA4DNN_CHECK_CUDA(cub::DeviceSegmentedSort::StableSortPairs(
+                    nullptr, temp_bytes, static_cast<const T*>(nullptr), static_cast<T*>(nullptr),
+                    static_cast<const int*>(nullptr), static_cast<int*>(nullptr), n, outer,
+                    static_cast<const int*>(nullptr), static_cast<const int*>(nullptr), stream));
+
+            TopKSortLayout layout;
+            layout.keys = 0;
+            layout.ids_in = layout.keys + alignTo256(static_cast<std::size_t>(n) * sizeof(T));
+            layout.ids_out = layout.ids_in + alignTo256(static_cast<std::size_t>(n) * sizeof(int));
+            layout.offsets = layout.ids_out + alignTo256(static_cast<std::size_t>(n) * sizeof(int));
+            layout.temp = layout.offsets + alignTo256(static_cast<std::size_t>(outer + 1) * sizeof(int));
+            layout.total = layout.temp + alignTo256(temp_bytes);
+            return layout;
+        }
+    }
+
+    template <class T>
+    void topk(const Stream& stream,
+        Span<T> values, Span<std::int64_t> indices, View<T> input,
+        int outer, int dim_axis, int inner, int k, bool largest)
+    {
+        CV_Assert(outer > 0 && dim_axis > 0 && inner > 0);
+        CV_Assert(k > 0 && k <= dim_axis);
+
+        constexpr int block_size = 256;
+        constexpr int max_blocks = 65535;
+        const int nslices = outer * inner;
+        const int grid_size = std::min(max_blocks, nslices);
+
+        raw::topk<T, block_size><<<grid_size, block_size, 0, stream.get()>>>(
+            values, indices, input, outer, dim_axis, inner, k, largest);
+    }
+
+    template <class T>
+    std::size_t topk_sort_workspace(const Stream& stream, int outer, int dim_axis, bool largest)
+    {
+        return detail::topkSortLayout<T>(outer, dim_axis, largest, stream.get()).total;
+    }
+
+    template <class T>
+    void topk_sort(const Stream& stream,
+        Span<T> values, Span<std::int64_t> indices, View<T> input,
+        int outer, int dim_axis, int k, bool largest,
+        void* workspace, std::size_t workspace_bytes)
+    {
+        CV_Assert(outer > 0 && dim_axis > 0);
+        CV_Assert(k > 0 && k <= dim_axis);
+
+        const detail::TopKSortLayout layout = detail::topkSortLayout<T>(outer, dim_axis, largest, stream.get());
+        CV_Assert(workspace != nullptr && workspace_bytes >= layout.total);
+
+        unsigned char* base = static_cast<unsigned char*>(workspace);
+        T* keys = reinterpret_cast<T*>(base + layout.keys);
+        int* ids_in = reinterpret_cast<int*>(base + layout.ids_in);
+        int* ids_out = reinterpret_cast<int*>(base + layout.ids_out);
+        int* offsets = reinterpret_cast<int*>(base + layout.offsets);
+        void* temp = base + layout.temp;
+        std::size_t temp_bytes = layout.total - layout.temp;
+
+        const int n = outer * dim_axis;
+        constexpr int block_size = 256;
+        constexpr int max_blocks = 4096;
+
+        const int init_grid = std::min(max_blocks, (std::max(n, outer + 1) + block_size - 1) / block_size);
+        raw::topk_sort_init<<<init_grid, block_size, 0, stream.get()>>>(ids_in, offsets, n, outer, dim_axis);
+
+        const T* keys_in = input.data().get();
+        if (outer == 1 && largest)
+            CUDA4DNN_CHECK_CUDA(cub::DeviceRadixSort::SortPairsDescending(
+                temp, temp_bytes, keys_in, keys, ids_in, ids_out, n, 0, static_cast<int>(sizeof(T) * 8), stream.get()));
+        else if (outer == 1)
+            CUDA4DNN_CHECK_CUDA(cub::DeviceRadixSort::SortPairs(
+                temp, temp_bytes, keys_in, keys, ids_in, ids_out, n, 0, static_cast<int>(sizeof(T) * 8), stream.get()));
+        else if (largest)
+            CUDA4DNN_CHECK_CUDA(cub::DeviceSegmentedSort::StableSortPairsDescending(
+                temp, temp_bytes, keys_in, keys, ids_in, ids_out, n, outer, offsets, offsets + 1, stream.get()));
+        else
+            CUDA4DNN_CHECK_CUDA(cub::DeviceSegmentedSort::StableSortPairs(
+                temp, temp_bytes, keys_in, keys, ids_in, ids_out, n, outer, offsets, offsets + 1, stream.get()));
+
+        const int m = outer * k;
+        const int gather_grid = std::min(max_blocks, (m + block_size - 1) / block_size);
+        raw::topk_sort_gather<T><<<gather_grid, block_size, 0, stream.get()>>>(values, indices, keys, ids_out, dim_axis, k);
+    }
+
+#if !defined(__CUDA_ARCH__) || (__CUDA_ARCH__ >= 530)
+    template void topk(const Stream&, Span<__half>, Span<std::int64_t>, View<__half>, int, int, int, int, bool);
+#endif
+    template void topk(const Stream&, Span<float>, Span<std::int64_t>, View<float>, int, int, int, int, bool);
+    template void topk(const Stream&, Span<int8_t>, Span<std::int64_t>, View<int8_t>, int, int, int, int, bool);
+    template void topk(const Stream&, Span<uint8_t>, Span<std::int64_t>, View<uint8_t>, int, int, int, int, bool);
+    template void topk(const Stream&, Span<int32_t>, Span<std::int64_t>, View<int32_t>, int, int, int, int, bool);
+    template void topk(const Stream&, Span<int64_t>, Span<std::int64_t>, View<int64_t>, int, int, int, int, bool);
+
+#define CUDA4DNN_INSTANTIATE_TOPK_SORT(T) \
+    template std::size_t topk_sort_workspace<T>(const Stream&, int, int, bool); \
+    template void topk_sort<T>(const Stream&, Span<T>, Span<std::int64_t>, View<T>, int, int, int, bool, void*, std::size_t);
+
+#if !defined(__CUDA_ARCH__) || (__CUDA_ARCH__ >= 530)
+    CUDA4DNN_INSTANTIATE_TOPK_SORT(__half)
+#endif
+    CUDA4DNN_INSTANTIATE_TOPK_SORT(float)
+    CUDA4DNN_INSTANTIATE_TOPK_SORT(int8_t)
+    CUDA4DNN_INSTANTIATE_TOPK_SORT(uint8_t)
+    CUDA4DNN_INSTANTIATE_TOPK_SORT(int32_t)
+    CUDA4DNN_INSTANTIATE_TOPK_SORT(int64_t)
+#undef CUDA4DNN_INSTANTIATE_TOPK_SORT
+
+}}}} /* namespace cv::dnn::cuda4dnn::kernels */

modules/dnn/src/cuda/trilu.cu

@@ -0,0 +1,75 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+// Copyright (C) 2026, BigVision LLC, all rights reserved.
+// Third party copyrights are property of their respective owners.
+#include <cuda_runtime.h>
+#include <cuda_fp16.h>
+
+#include "types.hpp"
+#include "grid_stride_range.hpp"
+#include "math.hpp"
+#include "execution.hpp"
+
+#include "../cuda4dnn/csl/stream.hpp"
+#include "../cuda4dnn/csl/span.hpp"
+
+#include "../cuda4dnn/kernels/trilu.hpp"
+
+#include <opencv2/core.hpp>
+
+#include <algorithm>
+#include <cstdint>
+
+using namespace cv::dnn::cuda4dnn::csl;
+using namespace cv::dnn::cuda4dnn::csl::device;
+
+namespace cv { namespace dnn { namespace cuda4dnn { namespace kernels {
+
+    namespace raw {
+
+        template <class T>
+        __global__ void trilu(Span<T> output, View<T> input,
+                              int height, int width, int k, bool upper)
+        {
+            const T* in = input.data().get();
+            T* out = output.data().get();
+
+            for (auto id : grid_stride_range(output.size())) {
+                const int c = id % width;
+                const int l = (id / width) % height;
+
+                // ONNX keeps c - l >= k (upper) or c - l <= k (lower) on every row, tall matrices included
+                const bool zero = upper ? (c - l < k) : (c - l > k);
+
+                out[id] = zero ? static_cast<T>(0.f) : in[id];
+            }
+        }
+    }
+
+    template <class T>
+    void trilu(const Stream& stream,
+        Span<T> output, View<T> input,
+        int loops, int height, int width, int k, bool upper)
+    {
+        CV_Assert(loops > 0 && height > 0 && width > 0);
+        CV_UNUSED(loops);
+
+        if (output.size() == 0)
+            return;
+
+        auto kernel = raw::trilu<T>;
+        auto policy = make_policy(kernel, output.size(), 0, stream);
+        launch_kernel(kernel, policy, output, input, height, width, k, upper);
+    }
+
+#if !defined(__CUDA_ARCH__) || (__CUDA_ARCH__ >= 530)
+    template void trilu(const Stream&, Span<__half>, View<__half>, int, int, int, int, bool);
+#endif
+    template void trilu(const Stream&, Span<float>, View<float>, int, int, int, int, bool);
+    template void trilu(const Stream&, Span<int32_t>, View<int32_t>, int, int, int, int, bool);
+    template void trilu(const Stream&, Span<int64_t>, View<int64_t>, int, int, int, int, bool);
+    template void trilu(const Stream&, Span<int8_t>, View<int8_t>, int, int, int, int, bool);
+    template void trilu(const Stream&, Span<uint8_t>, View<uint8_t>, int, int, int, int, bool);
+
+}}}} /* namespace cv::dnn::cuda4dnn::kernels */

modules/dnn/src/cuda4dnn/csl/cublas.hpp

@@ -11,7 +11,6 @@
 #include "memory.hpp"
 
 #include <opencv2/core.hpp>
-#include <opencv2/core/utils/configuration.private.hpp>
 
 #include <cublas_v2.h>
 

modules/dnn/src/cuda4dnn/csl/cudnn/graph.hpp

@@ -180,7 +180,10 @@ namespace cv { namespace dnn { namespace cuda4dnn { namespace csl { namespace cu
             return false;
         };
 
-        bool found = pick(CUDNN_HEUR_MODE_A, strictFP32) || (strictFP32 && pick(CUDNN_HEUR_MODE_FALLBACK, true));
+        // mode B ranks engines more accurately than mode A, still deterministically; if it does not
+        // finalize (e.g. its heuristics library is not loaded), pick() returns false and mode A decides
+        bool found = pick(CUDNN_HEUR_MODE_B, strictFP32) || pick(CUDNN_HEUR_MODE_A, strictFP32) ||
+                     (strictFP32 && pick(CUDNN_HEUR_MODE_FALLBACK, true));
         if (!found && strictFP32 && pick(CUDNN_HEUR_MODE_A, false))
             CV_LOG_ONCE_WARNING(NULL, "DNN/CUDA: cuDNN JIT has no exact FP32 engine for some " << what << " layers, using TF32 for them");
 

modules/dnn/src/cuda4dnn/kernels/argmax.hpp

@@ -0,0 +1,23 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+// Copyright (C) 2026, BigVision LLC, all rights reserved.
+// Third party copyrights are property of their respective owners.
+#ifndef OPENCV_DNN_SRC_CUDA4DNN_KERNELS_ARGMAX_HPP
+#define OPENCV_DNN_SRC_CUDA4DNN_KERNELS_ARGMAX_HPP
+
+#include "../csl/stream.hpp"
+#include "../csl/span.hpp"
+
+#include <cstdint>
+
+namespace cv { namespace dnn { namespace cuda4dnn { namespace kernels {
+
+template <class T>
+void arg_min_max(const csl::Stream& stream,
+    csl::Span<std::int64_t> output, csl::View<T> input,
+    int outer_size, int axis_size, int inner_size, bool is_argmax, bool select_last_index);
+
+}}}} /* namespace cv::dnn::cuda4dnn::kernels */
+
+#endif /* OPENCV_DNN_SRC_CUDA4DNN_KERNELS_ARGMAX_HPP */

modules/dnn/src/cuda4dnn/kernels/col2im.hpp

@@ -0,0 +1,29 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+// Copyright (C) 2026, BigVision LLC, all rights reserved.
+// Third party copyrights are property of their respective owners.
+
+#ifndef OPENCV_DNN_SRC_CUDA4DNN_KERNELS_COL2IM_HPP
+#define OPENCV_DNN_SRC_CUDA4DNN_KERNELS_COL2IM_HPP
+
+#include "../csl/stream.hpp"
+#include "../csl/span.hpp"
+
+namespace cv { namespace dnn { namespace cuda4dnn { namespace kernels {
+
+    struct Col2ImParams {
+        int batch, channels, total_channels, channel_offset;
+        int in_h, in_w, out_h, out_w;
+        int kernel_h, kernel_w, stride_h, stride_w, pad_h, pad_w, dil_h, dil_w;
+    };
+
+    // cols is [batch][channels * kernel_h * kernel_w][in_h * in_w]; each output pixel gathers the
+    // column entries that land on it, plus bias[channel_offset + c] when bias is not empty
+    template <class T>
+    void col2im_2d(const csl::Stream& stream, csl::Span<T> output, csl::View<T> cols, csl::View<T> bias,
+                   const Col2ImParams& params);
+
+}}}} /* namespace cv::dnn::cuda4dnn::kernels */
+
+#endif /* OPENCV_DNN_SRC_CUDA4DNN_KERNELS_COL2IM_HPP */

modules/dnn/src/cuda4dnn/kernels/cum_scan.hpp

@@ -0,0 +1,35 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+// Copyright (C) 2026, BigVision LLC, all rights reserved.
+// Third party copyrights are property of their respective owners.
+#ifndef OPENCV_DNN_SRC_CUDA4DNN_KERNELS_CUM_SCAN_HPP
+#define OPENCV_DNN_SRC_CUDA4DNN_KERNELS_CUM_SCAN_HPP
+
+#include "../csl/stream.hpp"
+#include "../csl/span.hpp"
+
+#include <cstddef>
+
+namespace cv { namespace dnn { namespace cuda4dnn { namespace kernels {
+
+/* ONNX CumSum / CumProd over one axis, with the input viewed as
+ * [outer_size, target_size, inner_size].
+ *
+ * `is_prod` picks the operator and the exclusive-mode seed (0 for sum, 1 for product);
+ * `exclusive` and `reverse` follow the ONNX attributes. Mirrors the scan in
+ * cumsum_layer.cpp / cumprod_layer.cpp exactly, including their offset arithmetic.
+ *
+ * The scan along the axis is inherently sequential, so a thread owns one
+ * (outer, inner) pair and walks the axis; parallelism comes from outer * inner, which
+ * is where the width is in practice.
+ */
+template <class T>
+void cum_scan(const csl::Stream& stream,
+    csl::Span<T> output, csl::View<T> input,
+    std::size_t outer_size, std::size_t target_size, std::size_t inner_size,
+    bool is_prod, bool exclusive, bool reverse);
+
+}}}} /* namespace cv::dnn::cuda4dnn::kernels */
+
+#endif /* OPENCV_DNN_SRC_CUDA4DNN_KERNELS_CUM_SCAN_HPP */

modules/dnn/src/cuda4dnn/kernels/eltwise_ops.hpp

@@ -42,6 +42,19 @@ namespace cv { namespace dnn { namespace cuda4dnn { namespace kernels {
     template <class T>
     void eltwise_pow_2(const csl::Stream& stream, csl::TensorSpan<T> output, csl::TensorView<T> x, csl::TensorView<T> y);
 
+    template <class T>
+    void eltwise_prelu_2(const csl::Stream& stream, csl::TensorSpan<T> output, csl::TensorView<T> x, csl::TensorView<T> slope);
+
+    // integer types only
+    template <class T>
+    void eltwise_bitand_2(const csl::Stream& stream, csl::TensorSpan<T> output, csl::TensorView<T> x, csl::TensorView<T> y);
+
+    template <class T>
+    void eltwise_bitor_2(const csl::Stream& stream, csl::TensorSpan<T> output, csl::TensorView<T> x, csl::TensorView<T> y);
+
+    template <class T>
+    void eltwise_bitxor_2(const csl::Stream& stream, csl::TensorSpan<T> output, csl::TensorView<T> x, csl::TensorView<T> y);
+
 }}}} /* namespace cv::dnn::cuda4dnn::kernels */
 
 #endif /* OPENCV_DNN_SRC_CUDA4DNN_KERNELS_ELTWISE_OPS_HPP */

modules/dnn/src/cuda4dnn/kernels/gather_elements.hpp

@@ -0,0 +1,28 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+// Copyright (C) 2026, BigVision LLC, all rights reserved.
+// Third party copyrights are property of their respective owners.
+#ifndef OPENCV_DNN_SRC_CUDA4DNN_KERNELS_GATHER_ELEMENTS_HPP
+#define OPENCV_DNN_SRC_CUDA4DNN_KERNELS_GATHER_ELEMENTS_HPP
+
+#include "../csl/stream.hpp"
+#include "../csl/span.hpp"
+
+#include "../../cuda/fast_divmod.hpp"
+
+#include <cstddef>
+#include <cstdint>
+
+namespace cv { namespace dnn { namespace cuda4dnn { namespace kernels {
+
+template <class T>
+void gather_elements(const csl::Stream& stream,
+    csl::Span<T> output, csl::View<T> input,
+    const void* indices, std::size_t index_element_size,
+    std::int64_t axis_size_data, std::int64_t inner_size,
+    csl::device::fast_divmod axis_inner_size, csl::device::fast_divmod inner_size_divmod);
+
+}}}} /* namespace cv::dnn::cuda4dnn::kernels */
+
+#endif /* OPENCV_DNN_SRC_CUDA4DNN_KERNELS_GATHER_ELEMENTS_HPP */

modules/dnn/src/cuda4dnn/kernels/gather_nd.hpp

@@ -0,0 +1,30 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+// Copyright (C) 2026, BigVision LLC, all rights reserved.
+// Third party copyrights are property of their respective owners.
+#ifndef OPENCV_DNN_SRC_CUDA4DNN_KERNELS_GATHER_ND_HPP
+#define OPENCV_DNN_SRC_CUDA4DNN_KERNELS_GATHER_ND_HPP
+
+#include "../csl/stream.hpp"
+#include "../csl/span.hpp"
+
+#include <cstddef>
+#include <cstdint>
+#include <vector>
+
+namespace cv { namespace dnn { namespace cuda4dnn { namespace kernels {
+
+template <class T, class TIdx>
+void gather_nd(const csl::Stream& stream,
+    csl::Span<T> output, csl::View<TIdx> indices, csl::View<T> input,
+    std::int64_t last_index_dimension,
+    const std::vector<std::int64_t>& element_counts,
+    const std::vector<std::int64_t>& input_dims,
+    std::int64_t inner_size,
+    std::int64_t batch_stride,
+    std::int64_t num_slices_per_batch);
+
+}}}} /* namespace cv::dnn::cuda4dnn::kernels */
+
+#endif /* OPENCV_DNN_SRC_CUDA4DNN_KERNELS_GATHER_ND_HPP */

modules/dnn/src/cuda4dnn/kernels/grid_sample.hpp

@@ -0,0 +1,37 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+// Copyright (C) 2026, BigVision LLC, all rights reserved.
+// Third party copyrights are property of their respective owners.
+#ifndef OPENCV_DNN_SRC_CUDA4DNN_KERNELS_GRID_SAMPLE_HPP
+#define OPENCV_DNN_SRC_CUDA4DNN_KERNELS_GRID_SAMPLE_HPP
+
+#include "../csl/stream.hpp"
+#include "../csl/span.hpp"
+
+namespace cv { namespace dnn { namespace cuda4dnn { namespace kernels {
+
+struct GridSampleParams {
+    int channels{}, height{}, width{};
+    int output_height{}, output_width{};
+    int mode{1};      // 0 = nearest, 1 = bilinear, 2 = bicubic
+    int padding{0};   // 0 = zeros,   1 = border,   2 = reflection
+    bool align_corners{false};
+    float cubic_alpha{-0.75f};
+};
+
+/* ONNX GridSample over 4-D input. `input` is [N, C, H, W], `grid` is
+ * [N, H_out, W_out, 2] in CV_32F holding normalised (x, y), and the output is
+ * [N, C, H_out, W_out].
+ *
+ * Mirrors gridSampleComputeRows() in gridsample_layer.cpp. That function has a fast
+ * path for fully-interior taps and a general `fetch` path; the two agree by
+ * construction, so this kernel only implements the general one. */
+template <class T>
+void grid_sample(const csl::Stream& stream,
+    csl::Span<T> output, csl::View<T> input, csl::View<float> grid,
+    const GridSampleParams& params);
+
+}}}} /* namespace cv::dnn::cuda4dnn::kernels */
+
+#endif /* OPENCV_DNN_SRC_CUDA4DNN_KERNELS_GRID_SAMPLE_HPP */

modules/dnn/src/cuda4dnn/kernels/mvn.hpp

@@ -18,6 +18,14 @@ void reduce_mean(const csl::Stream& stream, csl::Span<float> means, csl::View<T>
 template <class T>
 void reduce_mean_sqr_sum(const csl::Stream& stream, csl::Span<float> means, csl::Span<float> sum_sqrs, csl::View<T> input, std::size_t inner_size);
 
+// floats of scratch the overload below needs to split rows this long across blocks; 0 when it does not split
+std::size_t reduce_mean_sqr_sum_scratch(std::size_t rows, std::size_t inner_size);
+
+// same result, with long rows split across blocks and the partials added in a fixed order
+template <class T>
+void reduce_mean_sqr_sum(const csl::Stream& stream, csl::Span<float> means, csl::Span<float> sum_sqrs, csl::View<T> input,
+                         std::size_t inner_size, csl::Span<float> scratch);
+
 void compute_normalization_scale(const csl::Stream& stream, csl::Span<float> scale, csl::View<float> means, csl::View<float> sum_sqrs, std::size_t inner_size, float eps);
 
 template <class T>

modules/dnn/src/cuda4dnn/kernels/normalize.hpp

@@ -19,6 +19,14 @@ namespace cv { namespace dnn { namespace cuda4dnn { namespace kernels {
         std::size_t outer_size, std::size_t mid_size, std::size_t inner_size, std::size_t norm, T epsilon,
         csl::Span<T> workspace);
 
+    // ONNX LRN: y = x * (bias + alpha_scaled * sum of x^2 over channels [c - radius, c + radius])^-beta
+    template <class T>
+    void lrn_across_channels(
+        const csl::Stream& stream,
+        csl::Span<T> output, csl::View<T> input,
+        std::size_t batch, std::size_t channels, std::size_t plane_size,
+        int radius, float alpha_scaled, float bias, float beta);
+
 }}}} /* namespace cv::dnn::cuda4dnn::kernels */
 
 #endif /* OPENCV_DNN_SRC_CUDA4DNN_KERNELS_NORMALIZE_HPP */

modules/dnn/src/cuda4dnn/kernels/one_hot.hpp

@@ -0,0 +1,22 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+// Copyright (C) 2026, BigVision LLC, all rights reserved.
+// Third party copyrights are property of their respective owners.
+#ifndef OPENCV_DNN_SRC_CUDA4DNN_KERNELS_ONE_HOT_HPP
+#define OPENCV_DNN_SRC_CUDA4DNN_KERNELS_ONE_HOT_HPP
+
+#include "../csl/stream.hpp"
+#include "../csl/span.hpp"
+
+#include <cstddef>
+
+namespace cv { namespace dnn { namespace cuda4dnn { namespace kernels {
+
+    template <class T, class TIdx>
+    void one_hot(const csl::Stream& stream, csl::Span<T> output, csl::View<TIdx> indices,
+                 std::size_t depth, std::size_t inner, T off_value, T on_value);
+
+}}}} /* namespace cv::dnn::cuda4dnn::kernels */
+
+#endif /* OPENCV_DNN_SRC_CUDA4DNN_KERNELS_ONE_HOT_HPP */

modules/dnn/src/cuda4dnn/kernels/padding.hpp

@@ -20,6 +20,16 @@ namespace cv { namespace dnn { namespace cuda4dnn { namespace kernels {
         csl::TensorSpan<T> output, csl::TensorView<T> input,
         std::vector<std::pair<std::size_t, std::size_t>> ranges);
 
+    // border rules follow cv::borderInterpolate, as the CPU Pad2 does
+    enum class PadMode { CONSTANT, REFLECT101, REPLICATE, WRAP };
+
+    // pads_begin may be negative (crop); output shape already includes the pads
+    template <class T>
+    void pad_nd(
+        const csl::Stream& stream,
+        csl::TensorSpan<T> output, csl::TensorView<T> input,
+        const std::vector<int>& pads_begin, PadMode mode, T value);
+
 }}}} /* namespace cv::dnn::cuda4dnn::kernels */
 
 #endif /* OPENCV_DNN_SRC_CUDA4DNN_KERNELS_PADDING_HPP */

modules/dnn/src/cuda4dnn/kernels/quantize_dequantize.hpp

@@ -0,0 +1,38 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+
+#ifndef OPENCV_DNN_SRC_CUDA4DNN_KERNELS_QUANTIZE_DEQUANTIZE_HPP
+#define OPENCV_DNN_SRC_CUDA4DNN_KERNELS_QUANTIZE_DEQUANTIZE_HPP
+
+#include "../csl/stream.hpp"
+#include "../csl/span.hpp"
+
+#include <cstddef>
+#include <cstdint>
+
+namespace cv { namespace dnn { namespace cuda4dnn { namespace kernels {
+
+// Per-tensor (sz_a == 1) and per-axis (sz_a == input.size(axis)) ONNX DequantizeLinear/
+// QuantizeLinear, block_size == 0 only. zero_point may be an empty view (implies zero).
+template <class T>
+void dequantize_linear(const csl::Stream& stream, csl::Span<float> output, csl::View<T> input,
+                        csl::View<float> scale, csl::View<T> zero_point,
+                        std::size_t sz_a, std::size_t slice_size);
+
+template <class T>
+void quantize_linear(const csl::Stream& stream, csl::Span<T> output, csl::View<float> input,
+                      csl::View<float> scale, csl::View<T> zero_point,
+                      std::size_t sz_a, std::size_t slice_size);
+
+// Number of (min, max) partial pairs dynamic_quantize_scale() keeps in scratch for n inputs.
+std::size_t dynamic_quantize_partials(std::size_t n);
+
+// ONNX DynamicQuantizeLinear y_scale and uint8 y_zero_point from the input range, which is forced
+// to include zero; scratch holds 2 * dynamic_quantize_partials(input.size()) floats.
+void dynamic_quantize_scale(const csl::Stream& stream, csl::Span<float> scale, csl::Span<std::uint8_t> zero_point,
+                            csl::View<float> input, csl::Span<float> scratch);
+
+}}}} /* namespace cv::dnn::cuda4dnn::kernels */
+
+#endif /* OPENCV_DNN_SRC_CUDA4DNN_KERNELS_QUANTIZE_DEQUANTIZE_HPP */

modules/dnn/src/cuda4dnn/kernels/range.hpp

@@ -0,0 +1,20 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+// Copyright (C) 2026, BigVision LLC, all rights reserved.
+// Third party copyrights are property of their respective owners.
+#ifndef OPENCV_DNN_SRC_CUDA4DNN_KERNELS_RANGE_HPP
+#define OPENCV_DNN_SRC_CUDA4DNN_KERNELS_RANGE_HPP
+
+#include "../csl/stream.hpp"
+#include "../csl/span.hpp"
+
+namespace cv { namespace dnn { namespace cuda4dnn { namespace kernels {
+
+/* start and delta are read on the device, so no host round trip is needed */
+template <class T>
+void range(const csl::Stream& stream, csl::Span<T> output, const T* start, const T* delta);
+
+}}}} /* namespace cv::dnn::cuda4dnn::kernels */
+
+#endif /* OPENCV_DNN_SRC_CUDA4DNN_KERNELS_RANGE_HPP */

modules/dnn/src/cuda4dnn/kernels/resize.hpp

@@ -7,9 +7,18 @@
 
 #include "../csl/stream.hpp"
 #include "../csl/tensor.hpp"
+#include "../csl/span.hpp"
 
 namespace cv { namespace dnn { namespace cuda4dnn { namespace kernels {
 
+    // separable gather over host-built per-axis taps (idx < 0 skips a tap); a pixel whose row or
+    // column is flagged out of bounds takes extrapolation_value (tf_crop_and_resize)
+    template <class T>
+    void resize_separable(const csl::Stream& stream, csl::TensorSpan<T> output, csl::TensorView<T> input,
+                          csl::View<int> idx_y, csl::View<float> w_y, csl::View<int> oob_y, int taps_y,
+                          csl::View<int> idx_x, csl::View<float> w_x, csl::View<int> oob_x, int taps_x,
+                          float extrapolation_value);
+
     template <class T>
     void resize_nn(const csl::Stream& stream, csl::TensorSpan<T> output, csl::TensorView<T> input, float scale_y, float scale_x, bool round, bool half_pixel_centers);
 

modules/dnn/src/cuda4dnn/kernels/rms_norm.hpp

@@ -0,0 +1,34 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+// Copyright (C) 2026, BigVision LLC, all rights reserved.
+// Third party copyrights are property of their respective owners.
+#ifndef OPENCV_DNN_SRC_CUDA4DNN_KERNELS_RMS_NORM_HPP
+#define OPENCV_DNN_SRC_CUDA4DNN_KERNELS_RMS_NORM_HPP
+
+#include "../csl/stream.hpp"
+#include "../csl/span.hpp"
+
+#include <cstddef>
+
+namespace cv { namespace dnn { namespace cuda4dnn { namespace kernels {
+
+/* RMS normalisation: y = scale * x / sqrt(mean(x^2) + epsilon), applied over the last
+ * `norm_size` elements of each of `loops` rows. `scale` holds `norm_size` elements.
+ *
+ * Mirrors fastNormImpl(input, scale, output, eps, axis, recenter=false) in
+ * cpu_kernels/fast_norm.cpp -- the recenter=false path forces mean to 0, so the
+ * variance term collapses to the mean square.
+ *
+ * Unlike the LayerNorm path this needs no workspace: one block owns one row and
+ * reduces it in registers plus shared memory, instead of accumulating into a global
+ * scratch buffer with atomicAdd.
+ */
+template <class T>
+void rms_norm(const csl::Stream& stream,
+    csl::Span<T> output, csl::View<T> input, csl::View<T> scale,
+    std::size_t loops, std::size_t norm_size, float epsilon);
+
+}}}} /* namespace cv::dnn::cuda4dnn::kernels */
+
+#endif /* OPENCV_DNN_SRC_CUDA4DNN_KERNELS_RMS_NORM_HPP */

modules/dnn/src/cuda4dnn/kernels/roi_align.hpp

@@ -0,0 +1,44 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+// Copyright (C) 2026, BigVision LLC, all rights reserved.
+// Third party copyrights are property of their respective owners.
+#ifndef OPENCV_DNN_SRC_CUDA4DNN_KERNELS_ROI_ALIGN_HPP
+#define OPENCV_DNN_SRC_CUDA4DNN_KERNELS_ROI_ALIGN_HPP
+
+#include "../csl/stream.hpp"
+#include "../csl/span.hpp"
+
+#include <cstddef>
+#include <cstdint>
+
+namespace cv { namespace dnn { namespace cuda4dnn { namespace kernels {
+
+struct RoiAlignParams {
+    int batch_size{}, channels{}, height{}, width{};
+    int output_height{}, output_width{};
+    int sampling_ratio{};
+    float spatial_scale{1.f};
+    /* half_pixel subtracts 0.5 from the scaled ROI corners; output_half_pixel does not
+     * but clamps a degenerate ROI up to one unit. They are the two ONNX
+     * coordinate_transformation_mode values and they move independently. */
+    float offset{0.f};
+    bool clamp_malformed_roi{false};
+    bool max_mode{false};
+};
+
+/* ONNX RoiAlign. `rois` is [num_rois, 4] as (x1, y1, x2, y2) in CV_32F; `batch_indices`
+ * is [num_rois] of either int32 or int64, selected by `batch_index_element_size`.
+ * Output is [num_rois, channels, output_height, output_width].
+ *
+ * Mirrors the CPU RoiAlignForwardInvoker element for element, including its rule that a
+ * sample falling fully outside [-1, extent] contributes nothing at all. */
+template <class T>
+void roi_align(const csl::Stream& stream,
+    csl::Span<T> output, csl::View<T> input, csl::View<float> rois,
+    const void* batch_indices, std::size_t batch_index_element_size,
+    const RoiAlignParams& params);
+
+}}}} /* namespace cv::dnn::cuda4dnn::kernels */
+
+#endif /* OPENCV_DNN_SRC_CUDA4DNN_KERNELS_ROI_ALIGN_HPP */

modules/dnn/src/cuda4dnn/kernels/scatter_elements.hpp

@@ -0,0 +1,27 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+// Copyright (C) 2026, BigVision LLC, all rights reserved.
+// Third party copyrights are property of their respective owners.
+#ifndef OPENCV_DNN_SRC_CUDA4DNN_KERNELS_SCATTER_ELEMENTS_HPP
+#define OPENCV_DNN_SRC_CUDA4DNN_KERNELS_SCATTER_ELEMENTS_HPP
+
+#include "../csl/stream.hpp"
+#include "../csl/span.hpp"
+
+#include <cstdint>
+#include <vector>
+
+namespace cv { namespace dnn { namespace cuda4dnn { namespace kernels {
+
+template <class T, class TIdx>
+void scatter_elements(const csl::Stream& stream,
+    csl::Span<T> output, csl::View<TIdx> indices, csl::View<T> updates,
+    int rank, int axis,
+    const std::vector<std::int64_t>& indices_dims,
+    const std::vector<std::int64_t>& output_strides,
+    std::int64_t axis_dim_size);
+
+}}}} /* namespace cv::dnn::cuda4dnn::kernels */
+
+#endif /* OPENCV_DNN_SRC_CUDA4DNN_KERNELS_SCATTER_ELEMENTS_HPP */

modules/dnn/src/cuda4dnn/kernels/scatter_nd.hpp

@@ -0,0 +1,28 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+// Copyright (C) 2026, BigVision LLC, all rights reserved.
+// Third party copyrights are property of their respective owners.
+#ifndef OPENCV_DNN_SRC_CUDA4DNN_KERNELS_SCATTER_ND_HPP
+#define OPENCV_DNN_SRC_CUDA4DNN_KERNELS_SCATTER_ND_HPP
+
+#include "../csl/stream.hpp"
+#include "../csl/span.hpp"
+
+#include <cstddef>
+#include <cstdint>
+#include <vector>
+
+namespace cv { namespace dnn { namespace cuda4dnn { namespace kernels {
+
+template <class T, class TIdx>
+void scatter_nd(const csl::Stream& stream,
+    csl::Span<T> output, csl::View<TIdx> indices, csl::View<T> updates,
+    std::size_t num_indices, std::int64_t last_index_dimension,
+    const std::vector<std::int64_t>& element_counts,
+    const std::vector<std::int64_t>& input_dims,
+    std::size_t num_updates_elements);
+
+}}}} /* namespace cv::dnn::cuda4dnn::kernels */
+
+#endif /* OPENCV_DNN_SRC_CUDA4DNN_KERNELS_SCATTER_ND_HPP */

modules/dnn/src/cuda4dnn/kernels/topk.hpp

@@ -0,0 +1,44 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+// Copyright (C) 2026, BigVision LLC, all rights reserved.
+// Third party copyrights are property of their respective owners.
+#ifndef OPENCV_DNN_SRC_CUDA4DNN_KERNELS_TOPK_HPP
+#define OPENCV_DNN_SRC_CUDA4DNN_KERNELS_TOPK_HPP
+
+#include "../csl/stream.hpp"
+#include "../csl/span.hpp"
+
+#include <cstddef>
+#include <cstdint>
+
+namespace cv { namespace dnn { namespace cuda4dnn { namespace kernels {
+
+/* ONNX TopK along one axis. The input is treated as [outer, dim_axis, inner] and the
+ * outputs as [outer, k, inner], matching TopK2's CPU layout.
+ *
+ * Selection is by strict lexicographic order -- value first, then smaller index on a
+ * tie -- which is what ComparatorGreater/ComparatorLess in topk2_layer.cpp impose. The
+ * results come out in rank order, so `sorted=1` is satisfied and `sorted=0` (which
+ * leaves order unspecified) is satisfied too.
+ *
+ * One block owns one slice and makes `k` passes over it, so the cost is O(k * dim_axis).
+ * That suits the small k detection heads use; it is not a general large-k sort.
+ */
+template <class T>
+void topk(const csl::Stream& stream,
+    csl::Span<T> values, csl::Span<std::int64_t> indices, csl::View<T> input,
+    int outer, int dim_axis, int inner, int k, bool largest);
+
+template <class T>
+std::size_t topk_sort_workspace(const csl::Stream& stream, int outer, int dim_axis, bool largest);
+
+template <class T>
+void topk_sort(const csl::Stream& stream,
+    csl::Span<T> values, csl::Span<std::int64_t> indices, csl::View<T> input,
+    int outer, int dim_axis, int k, bool largest,
+    void* workspace, std::size_t workspace_bytes);
+
+}}}} /* namespace cv::dnn::cuda4dnn::kernels */
+
+#endif /* OPENCV_DNN_SRC_CUDA4DNN_KERNELS_TOPK_HPP */

modules/dnn/src/cuda4dnn/kernels/trilu.hpp

@@ -0,0 +1,22 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+// Copyright (C) 2026, BigVision LLC, all rights reserved.
+// Third party copyrights are property of their respective owners.
+#ifndef OPENCV_DNN_SRC_CUDA4DNN_KERNELS_TRILU_HPP
+#define OPENCV_DNN_SRC_CUDA4DNN_KERNELS_TRILU_HPP
+
+#include "../csl/stream.hpp"
+#include "../csl/span.hpp"
+
+namespace cv { namespace dnn { namespace cuda4dnn { namespace kernels {
+
+// ONNX Trilu over the trailing [height, width] of each of `loops` matrices
+template <class T>
+void trilu(const csl::Stream& stream,
+    csl::Span<T> output, csl::View<T> input,
+    int loops, int height, int width, int k, bool upper);
+
+}}}} /* namespace cv::dnn::cuda4dnn::kernels */
+
+#endif /* OPENCV_DNN_SRC_CUDA4DNN_KERNELS_TRILU_HPP */

modules/dnn/src/cuda4dnn/primitives/activation.hpp

@@ -53,6 +53,37 @@ namespace cv { namespace dnn { namespace cuda4dnn {
         const T slope;
     };
 
+    // ONNX Clip: min/max may arrive as extra inputs, so only input 0 is data
+    template <class T>
+    class ClipOp final : public CUDABackendNode {
+    public:
+        ClipOp(csl::Stream stream_, double min_, double max_)
+            : stream(std::move(stream_)), min{ toT(min_) }, max{ toT(max_) } { }
+
+        void forward(
+            const std::vector<UMat>& inputs,
+            const std::vector<UMat>& outputs,
+            csl::Workspace& workspace) override
+        {
+            CV_UNUSED(workspace);
+            CV_Assert(!inputs.empty() && outputs.size() == 1);
+            auto input = csl::viewOf<T>(inputs[0]);
+            auto output = csl::spanOf<T>(outputs[0]);
+            kernels::clipped_relu<T>(stream, output, input, min, max);
+        }
+
+    private:
+        // integers convert straight from double so int64 bounds keep their precision; half goes through float
+        static T toT(double v)
+        {
+            using U = typename std::conditional<std::is_integral<T>::value, double, float>::type;
+            return static_cast<T>(static_cast<U>(v));
+        }
+
+        csl::Stream stream;
+        const T min, max;
+    };
+
     template <class T>
     class ClippedReLUOp final : public BaseOp<ClippedReLUOp, T> {
     public:

modules/dnn/src/cuda4dnn/primitives/arg.hpp

@@ -0,0 +1,65 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+// Copyright (C) 2026, BigVision LLC, all rights reserved.
+// Third party copyrights are property of their respective owners.
+#ifndef OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_ARG_HPP
+#define OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_ARG_HPP
+
+#include "../../op_cuda.hpp"
+
+#include "../csl/stream.hpp"
+
+#include "../kernels/argmax.hpp"
+
+#include <opencv2/core.hpp>
+
+#include <cstdint>
+#include <utility>
+
+namespace cv { namespace dnn { namespace cuda4dnn {
+
+    template <class T>
+    class ArgOp final : public CUDABackendNode {
+    public:
+        ArgOp(csl::Stream stream_, int axis_, bool is_argmax_, bool select_last_index_)
+            : stream(std::move(stream_)), axis(axis_), is_argmax(is_argmax_), select_last_index(select_last_index_)
+        {
+        }
+
+        void forward(
+            const std::vector<UMat>& inputs,
+            const std::vector<UMat>& outputs,
+            csl::Workspace& workspace) override
+        {
+            CV_UNUSED(workspace);
+            CV_Assert(inputs.size() == 1 && outputs.size() == 1);
+
+            auto input = csl::viewOf<T>(inputs[0]);
+            auto output = csl::spanOf<std::int64_t>(outputs[0]);
+
+            MatShape inShape = cv::dnn::shape(inputs[0]);
+            const int rank = inShape.dims;
+            const int ax = normalize_axis(axis, rank);
+
+            const int axis_size = inShape[ax];
+            int outer_size = 1;
+            for (int i = 0; i < ax; i++)
+                outer_size *= inShape[i];
+            int inner_size = 1;
+            for (int i = ax + 1; i < rank; i++)
+                inner_size *= inShape[i];
+
+            kernels::arg_min_max<T>(stream, output, input, outer_size, axis_size, inner_size, is_argmax, select_last_index);
+        }
+
+    private:
+        csl::Stream stream;
+        int axis;
+        bool is_argmax;
+        bool select_last_index;
+    };
+
+}}} /* namespace cv::dnn::cuda4dnn */
+
+#endif /* OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_ARG_HPP */

modules/dnn/src/cuda4dnn/primitives/average_pooling.hpp

@@ -12,6 +12,7 @@
 #include "../csl/tensor.hpp"
 
 #include "../kernels/average_pooling.hpp"
+#include "../kernels/reduce.hpp"
 
 #include <opencv2/core.hpp>
 
@@ -58,11 +59,35 @@ namespace cv { namespace dnn { namespace cuda4dnn {
             std::vector<std::int64_t> in_shape(inShape.begin(), inShape.end());
             std::vector<std::int64_t> out_shape(outShape.begin(), outShape.end());
 
+            // a window over every spatial axis without padding is a per-channel mean, which the
+            // reduction kernel spreads over a block per channel instead of one thread per output
+            if (isGlobal(in_shape)) {
+                std::vector<std::int64_t> axes;
+                for (std::size_t k = 2; k < in_shape.size(); k++)
+                    axes.push_back(static_cast<std::int64_t>(k));
+                kernels::reduce_mean<T>(stream, output, input, in_shape, axes);
+                return;
+            }
+
             kernels::average_pool<T>(stream, output, input,
                                      in_shape, out_shape, kernel_shape, strides, pads, dilations, count_include_pad);
         }
 
     private:
+        bool isGlobal(const std::vector<std::int64_t>& in_shape) const
+        {
+            const std::size_t nspatial = kernel_shape.size();
+            if (nspatial == 0 || in_shape.size() != nspatial + 2)
+                return false;
+            for (std::size_t i = 0; i < nspatial; i++)
+                if (kernel_shape[i] != in_shape[2 + i] || (!dilations.empty() && dilations[i] != 1))
+                    return false;
+            for (std::int64_t p : pads)
+                if (p != 0)
+                    return false;
+            return true;
+        }
+
         csl::Stream stream;
         std::vector<std::int64_t> kernel_shape, strides, pads, dilations;
         bool count_include_pad;

modules/dnn/src/cuda4dnn/primitives/conv_transpose2.hpp

@@ -0,0 +1,150 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+// Copyright (C) 2026, BigVision LLC, all rights reserved.
+// Third party copyrights are property of their respective owners.
+
+#ifndef OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_CONV_TRANSPOSE2_HPP
+#define OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_CONV_TRANSPOSE2_HPP
+
+#include "../../op_cuda.hpp"
+
+#include "../csl/stream.hpp"
+#include "../csl/cublas.hpp"
+#include "../csl/span.hpp"
+#include "../csl/tensor.hpp"
+#include "../csl/workspace.hpp"
+
+#include "../kernels/col2im.hpp"
+
+#include <opencv2/core.hpp>
+
+#include <cstddef>
+#include <vector>
+#include <utility>
+
+namespace cv { namespace dnn { namespace cuda4dnn {
+
+    struct ConvTranspose2Configuration {
+        MatShape input_shape, weight_shape, output_shape;  // NCHW, [Cin, Cout/group, kh, kw], NCHW
+        int groups;
+        std::vector<int> strides, dilations, pads_begin;   // two spatial entries each
+    };
+
+    // ConvTranspose without legacy cuDNN: per group, cols = W^T * X (cuBLAS), then col2im adds into Y
+    template <class T>
+    class ConvTranspose2Op final : public CUDABackendNode {
+    public:
+        ConvTranspose2Op(csl::Stream stream_, csl::cublas::Handle handle, const ConvTranspose2Configuration& config)
+            : stream(std::move(stream_)), cublasHandle(std::move(handle))
+        {
+            const MatShape& in = config.input_shape;
+            const MatShape& w = config.weight_shape;
+            const MatShape& out = config.output_shape;
+            CV_Assert(in.dims == 4 && w.dims == 4 && out.dims == 4);
+            CV_Assert(config.groups > 0 && in[1] % config.groups == 0 && w[0] == in[1]);
+
+            groups = config.groups;
+            batch = in[0];
+            cin = in[1];
+            cin_g = cin / groups;
+            cout_g = w[1];
+            cout = out[1];
+            CV_Assert(cout == cout_g * groups);
+            in_h = in[2]; in_w = in[3];
+            out_h = out[2]; out_w = out[3];
+            kernel_h = w[2]; kernel_w = w[3];
+            strides = config.strides;
+            dilations = config.dilations;
+            pads_begin = config.pads_begin;
+
+            // one group's columns for the whole batch; groups reuse it in stream order
+            csl::WorkspaceBuilder builder;
+            builder.require<T>(colsSize());
+            scratch_mem_in_bytes = builder.required_workspace_size();
+        }
+
+        // constant weights and bias the graph packed into the layer, uploaded once and owned by the node
+        ConvTranspose2Op(csl::Stream stream_, csl::cublas::Handle handle, const ConvTranspose2Configuration& config,
+                         const Mat& weights, const Mat& bias)
+            : ConvTranspose2Op(std::move(stream_), std::move(handle), config)
+        {
+            ownedWeights = upload(weights);
+            if (!bias.empty())
+                ownedBias = upload(bias);
+        }
+
+        void forward(
+            const std::vector<UMat>& inputs,
+            const std::vector<UMat>& outputs,
+            csl::Workspace& workspace) override
+        {
+            const bool owned = !ownedWeights.empty();
+            CV_Assert(inputs.size() >= (owned ? 1u : 2u) && outputs.size() == 1);
+
+            auto x = csl::viewOf<T>(inputs[0]);
+            auto w = csl::viewOf<T>(owned ? ownedWeights : inputs[1]);
+            csl::View<T> bias;
+            const UMat& biasMat = owned ? ownedBias : (inputs.size() >= 3 ? inputs[2] : UMat());
+            if (!biasMat.empty()) {
+                auto b = csl::viewOf<T>(biasMat);
+                bias = csl::View<T>(b.get(), b.size());
+            }
+            auto y = csl::spanOf<T>(outputs[0]);
+
+            auto ws_allocator = csl::WorkspaceAllocator(workspace);
+            auto cols = ws_allocator.get_span<T>(colsSize());
+
+            const std::size_t hw = (std::size_t)in_h * in_w;
+            const std::size_t m = (std::size_t)cout_g * kernel_h * kernel_w;
+            for (int g = 0; g < groups; g++) {
+                // row-major cols(m x hw) = W_g^T(m x cin_g) * X_g(cin_g x hw); in column-major terms cols^T = X_g^T * W_g
+                csl::cublas::gemmStridedBatched<T>(cublasHandle,
+                    false, true, hw, m, cin_g,
+                    T(1), x.get() + (std::ptrdiff_t)(g * cin_g * hw), hw, (std::size_t)cin * hw,
+                    w.get() + (std::ptrdiff_t)(g * cin_g * m), m, 0,
+                    T(0), cols.data(), hw, m * hw,
+                    batch);
+
+                kernels::Col2ImParams p;
+                p.batch = batch; p.channels = cout_g; p.total_channels = cout; p.channel_offset = g * cout_g;
+                p.in_h = in_h; p.in_w = in_w; p.out_h = out_h; p.out_w = out_w;
+                p.kernel_h = kernel_h; p.kernel_w = kernel_w;
+                p.stride_h = strides[0]; p.stride_w = strides[1];
+                p.pad_h = pads_begin[0]; p.pad_w = pads_begin[1];
+                p.dil_h = dilations[0]; p.dil_w = dilations[1];
+                kernels::col2im_2d<T>(stream, csl::Span<T>(y.get(), y.size()),
+                                      csl::View<T>(cols.data(), cols.size()), bias, p);
+            }
+        }
+
+        std::size_t get_workspace_memory_in_bytes() const noexcept override { return scratch_mem_in_bytes; }
+
+    private:
+        std::size_t colsSize() const { return (std::size_t)batch * cout_g * kernel_h * kernel_w * in_h * in_w; }
+
+        // FP32 host data -> device buffer in the node's element type
+        static UMat upload(const Mat& host32)
+        {
+            Mat host;
+            host32.convertTo(host, std::is_same<T, half>::value ? CV_16F : CV_32F);
+            UMat u;
+            u.allocator = cv::cuda::getCudaAllocator();
+            u.fitSameSize(host, host.type());
+            checkCudaAllocated(u);
+            host.copyTo(u);
+            return u;
+        }
+
+        UMat ownedWeights, ownedBias;
+        csl::Stream stream;
+        csl::cublas::Handle cublasHandle;
+        int groups, batch, cin, cin_g, cout, cout_g;
+        int in_h, in_w, out_h, out_w, kernel_h, kernel_w;
+        std::vector<int> strides, dilations, pads_begin;
+        std::size_t scratch_mem_in_bytes;
+    };
+
+}}} /* namespace cv::dnn::cuda4dnn */
+
+#endif /* OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_CONV_TRANSPOSE2_HPP */

modules/dnn/src/cuda4dnn/primitives/convolution.hpp

@@ -354,9 +354,12 @@ namespace cv { namespace dnn { namespace cuda4dnn {
                 builder.require<T>(sz);
             }
 #if defined(HAVE_CUDNNJIT) && !defined(HAVE_CUDNN)
-            // NHWC staging copies come from the shared workspace, taken before the conv scratchpad
-            builder.require<T>(std::accumulate(std::begin(jitInputShapeCL), std::end(jitInputShapeCL), std::size_t(1), std::multiplies<std::size_t>()));
-            builder.require<T>(std::accumulate(std::begin(jitOutputShapeCL), std::end(jitOutputShapeCL), std::size_t(1), std::multiplies<std::size_t>()));
+            if (!jitInputShapeCL.empty())
+            {
+                // NHWC staging copies come from the shared workspace, taken before the conv scratchpad
+                builder.require<T>(std::accumulate(std::begin(jitInputShapeCL), std::end(jitInputShapeCL), std::size_t(1), std::multiplies<std::size_t>()));
+                builder.require<T>(std::accumulate(std::begin(jitOutputShapeCL), std::end(jitOutputShapeCL), std::size_t(1), std::multiplies<std::size_t>()));
+            }
             builder.require(jitConvoluter.get_workspace_size());
 #else
             builder.require(convoluter.get_workspace_size());
@@ -386,8 +389,12 @@ namespace cv { namespace dnn { namespace cuda4dnn {
             }
 
 #if defined(HAVE_CUDNNJIT) && !defined(HAVE_CUDNN)
-            auto jitInputNHWC  = allocator.get_tensor_span<T>(std::begin(jitInputShapeCL), std::end(jitInputShapeCL));
-            auto jitOutputNHWC = allocator.get_tensor_span<T>(std::begin(jitOutputShapeCL), std::end(jitOutputShapeCL));
+            csl::TensorSpan<T> jitInputNHWC, jitOutputNHWC;
+            if (!jitInputShapeCL.empty())
+            {
+                jitInputNHWC  = allocator.get_tensor_span<T>(std::begin(jitInputShapeCL), std::end(jitInputShapeCL));
+                jitOutputNHWC = allocator.get_tensor_span<T>(std::begin(jitOutputShapeCL), std::end(jitOutputShapeCL));
+            }
 #endif
             auto conv_scratchpad = allocator.get_instance();
 

modules/dnn/src/cuda4dnn/primitives/cum_scan.hpp

@@ -0,0 +1,67 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+// Copyright (C) 2026, BigVision LLC, all rights reserved.
+// Third party copyrights are property of their respective owners.
+#ifndef OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_CUM_SCAN_HPP
+#define OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_CUM_SCAN_HPP
+
+#include "../../op_cuda.hpp"
+
+#include "../csl/stream.hpp"
+
+#include "../kernels/cum_scan.hpp"
+
+#include <opencv2/core.hpp>
+#include <opencv2/dnn/shape_utils.hpp>
+
+#include <cstddef>
+#include <utility>
+#include <vector>
+
+namespace cv { namespace dnn { namespace cuda4dnn {
+
+    /* Backs CumSum and CumProd. The axis is resolved by the layer at init time -- it can
+     * arrive as an input tensor, and reading it here would mean a device-to-host stall
+     * inside forward(). */
+    template <class T>
+    class CumScanOp final : public CUDABackendNode {
+    public:
+        CumScanOp(csl::Stream stream_, int axis_, bool is_prod_, bool exclusive_, bool reverse_)
+            : stream(std::move(stream_)), axis(axis_),
+              is_prod(is_prod_), exclusive(exclusive_), reverse(reverse_) { }
+
+        void forward(
+            const std::vector<UMat>& inputs,
+            const std::vector<UMat>& outputs,
+            csl::Workspace& workspace) override
+        {
+            CV_UNUSED(workspace);
+            CV_Assert(inputs.size() >= 1 && outputs.size() == 1);
+
+            auto input = csl::viewOf<T>(inputs[0]);
+            auto output = csl::spanOf<T>(outputs[0]);
+
+            MatShape inShape = cv::dnn::shape(inputs[0]);
+            const int a = normalize_axis(axis, inShape.dims);
+            CV_Assert(a >= 0 && a < inShape.dims);
+
+            std::size_t outer_size = 1, inner_size = 1;
+            for (int i = 0; i < a; i++) outer_size *= static_cast<std::size_t>(inShape[i]);
+            for (int i = a + 1; i < inShape.dims; i++) inner_size *= static_cast<std::size_t>(inShape[i]);
+            const std::size_t target_size = static_cast<std::size_t>(inShape[a]);
+
+            kernels::cum_scan<T>(stream, output, input,
+                                 outer_size, target_size, inner_size,
+                                 is_prod, exclusive, reverse);
+        }
+
+    private:
+        csl::Stream stream;
+        int axis;
+        bool is_prod, exclusive, reverse;
+    };
+
+}}} /* namespace cv::dnn::cuda4dnn */
+
+#endif /* OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_CUM_SCAN_HPP */

modules/dnn/src/cuda4dnn/primitives/depth_space_ops.hpp

@@ -40,17 +40,15 @@ namespace cv { namespace dnn { namespace cuda4dnn {
             scratch_mem_in_bytes = builder.required_workspace_size();
         }
 
-        void forward(const std::vector<cv::Ptr<BackendWrapper>> &inputs,
-                     const std::vector<cv::Ptr<BackendWrapper>> &outputs,
+        void forward(const std::vector<UMat> &inputs,
+                     const std::vector<UMat> &outputs,
                      csl::Workspace &workspace) override {
             CV_CheckEQ(inputs.size(), size_t(1), "DepthSpaceOps: only one input is accepted");
             CV_CheckEQ(outputs.size(), size_t(1), "DepthSpaceOps: only one output is accepted");
 
-            auto input_wrapper = inputs.front().dynamicCast<wrapper_type>();
-            auto input = input_wrapper->getView();
+            auto input = csl::viewOf<T>(inputs.front());
             CV_CheckEQ(input.rank(), size_t(4), "DepthSpaceOps: input needs to be 4-dimensional [N, C, H, W]");
-            auto output_wrapper = outputs.front().dynamicCast<wrapper_type>();
-            auto output = output_wrapper->getSpan();
+            auto output = csl::spanOf<T>(outputs.front());
             auto ws_allocator = csl::WorkspaceAllocator(workspace);
             auto transposed_internal = ws_allocator.get_tensor_span<T>(transposed_internal_shape.begin(), transposed_internal_shape.end());
 

modules/dnn/src/cuda4dnn/primitives/dropout_mask.hpp

@@ -0,0 +1,51 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+// Copyright (C) 2026, BigVision LLC, all rights reserved.
+// Third party copyrights are property of their respective owners.
+
+#ifndef OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_DROPOUT_MASK_HPP
+#define OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_DROPOUT_MASK_HPP
+
+#include "../../op_cuda.hpp"
+
+#include "../csl/stream.hpp"
+#include "../csl/tensor_ops.hpp"
+
+#include "../kernels/fill_copy.hpp"
+
+#include <vector>
+#include <utility>
+
+namespace cv { namespace dnn { namespace cuda4dnn {
+
+    // Inference-mode Dropout with its mask output(s): y = x, and every mask is all true.
+    template <class T>
+    class DropoutMaskOp final : public CUDABackendNode {
+    public:
+        DropoutMaskOp(csl::Stream stream_) : stream(std::move(stream_)) { }
+
+        void forward(
+            const std::vector<UMat>& inputs,
+            const std::vector<UMat>& outputs,
+            csl::Workspace& workspace) override
+        {
+            CV_UNUSED(workspace);
+            CV_Assert(!inputs.empty() && outputs.size() >= 2);
+
+            auto input = csl::viewOf<T>(inputs[0]);
+            auto output = csl::spanOf<T>(outputs[0]);
+            if (input.get() != output.get())
+                csl::tensor_ops::copy(stream, output, input);
+
+            for (std::size_t i = 1; i < outputs.size(); i++)
+                kernels::fill<bool>(stream, csl::spanOf<bool>(outputs[i]), true);
+        }
+
+    private:
+        csl::Stream stream;
+    };
+
+}}} /* namespace cv::dnn::cuda4dnn */
+
+#endif /* OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_DROPOUT_MASK_HPP */

modules/dnn/src/cuda4dnn/primitives/einsum.hpp

@@ -0,0 +1,108 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+// Copyright (C) 2026, BigVision LLC, all rights reserved.
+// Third party copyrights are property of their respective owners.
+
+#ifndef OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_EINSUM_HPP
+#define OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_EINSUM_HPP
+
+#include "../../op_cuda.hpp"
+
+#include "../csl/stream.hpp"
+#include "../csl/cublas.hpp"
+#include "../csl/tensor.hpp"
+#include "../csl/tensor_ops.hpp"
+#include "../csl/workspace.hpp"
+
+#include "../kernels/permute.hpp"
+
+#include <cstddef>
+#include <utility>
+#include <vector>
+
+namespace cv { namespace dnn { namespace cuda4dnn {
+
+    /** A two-operand einsum as permute -> strided batched GEMM -> permute.
+     *
+     * A is read as [batch, M, K] (or [batch, K, M] with transA) and B as [batch, K, N]
+     * (or [batch, N, K] with transB); the GEMM writes [batch, M, N]. An empty perm means
+     * the operand already has that layout, so it is used in place.
+     */
+    struct EinsumPlan {
+        std::vector<std::size_t> permA, permB, permC;     // permC maps [batch, M, N] to the output
+        std::vector<std::size_t> shapeA, shapeB, shapeC;  // full shapes after permA/permB, and of the GEMM result
+        bool transA = false, transB = false;
+        std::size_t batch = 1, M = 1, N = 1, K = 1;
+    };
+
+    template <class T>
+    class EinsumOp final : public CUDABackendNode {
+    public:
+        EinsumOp(csl::Stream stream_, csl::cublas::Handle handle_, EinsumPlan plan_)
+            : stream(std::move(stream_)), cublasHandle(std::move(handle_)), plan(std::move(plan_))
+        {
+            csl::WorkspaceBuilder builder;
+            if (!plan.permA.empty())
+                builder.require<T>(plan.batch * plan.M * plan.K);
+            if (!plan.permB.empty())
+                builder.require<T>(plan.batch * plan.K * plan.N);
+            if (!plan.permC.empty())
+                builder.require<T>(plan.batch * plan.M * plan.N);
+            scratch_mem_in_bytes = builder.required_workspace_size();
+        }
+
+        void forward(
+            const std::vector<UMat>& inputs,
+            const std::vector<UMat>& outputs,
+            csl::Workspace& workspace) override
+        {
+            CV_Assert(inputs.size() == 2 && outputs.size() == 1);
+
+            auto allocator = csl::WorkspaceAllocator(workspace);
+            csl::TensorView<T> a = operand(allocator, csl::viewOf<T>(inputs[0]), plan.permA, plan.shapeA);
+            csl::TensorView<T> b = operand(allocator, csl::viewOf<T>(inputs[1]), plan.permB, plan.shapeB);
+
+            const std::size_t rowsA = plan.transA ? plan.K : plan.M, colsA = plan.transA ? plan.M : plan.K;
+            const std::size_t rowsB = plan.transB ? plan.N : plan.K, colsB = plan.transB ? plan.K : plan.N;
+
+            auto output = csl::spanOf<T>(outputs[0]);
+            csl::TensorSpan<T> c = plan.permC.empty() ? output
+                : allocator.get_tensor_span<T>(plan.shapeC.begin(), plan.shapeC.end());
+
+            // reshape() cannot raise the rank, so the rank-3 GEMM views are built over the same memory
+            const std::size_t shapeA3[] = { plan.batch, rowsA, colsA };
+            const std::size_t shapeB3[] = { plan.batch, rowsB, colsB };
+            const std::size_t shapeC3[] = { plan.batch, plan.M, plan.N };
+            csl::TensorView<T> a3(a.get(), std::begin(shapeA3), std::end(shapeA3));
+            csl::TensorView<T> b3(b.get(), std::begin(shapeB3), std::end(shapeB3));
+            csl::TensorSpan<T> c3(c.get(), std::begin(shapeC3), std::end(shapeC3));
+            csl::tensor_ops::gemmStridedBatched<T>(cublasHandle, 0.0, c3, 1.0, plan.transA, a3, plan.transB, b3);
+
+            if (!plan.permC.empty())
+                kernels::permute<T>(stream, output, c, plan.permC);
+        }
+
+        std::size_t get_workspace_memory_in_bytes() const noexcept override { return scratch_mem_in_bytes; }
+
+    private:
+        // the operand itself, or a permuted copy in the workspace
+        csl::TensorView<T> operand(csl::WorkspaceAllocator& allocator, csl::TensorView<T> input,
+                                   const std::vector<std::size_t>& perm, const std::vector<std::size_t>& shape)
+        {
+            if (perm.empty())
+                return input;
+            auto permuted = allocator.get_tensor_span<T>(shape.begin(), shape.end());
+            kernels::permute<T>(stream, permuted, input, perm);
+            return permuted;
+        }
+
+        csl::Stream stream;
+        csl::cublas::Handle cublasHandle;
+        EinsumPlan plan;
+        std::size_t scratch_mem_in_bytes;
+    };
+
+}}} /* namespace cv::dnn::cuda4dnn */
+
+#endif /* OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_EINSUM_HPP */

modules/dnn/src/cuda4dnn/primitives/eltwise.hpp

@@ -16,6 +16,7 @@
 #include <opencv2/core.hpp>
 
 #include <cstddef>
+#include <type_traits>
 #include <vector>
 #include <utility>
 
@@ -31,6 +32,10 @@ namespace cv { namespace dnn { namespace cuda4dnn {
         MOD,
         FMOD,
         POW,
+        PRELU,
+        BIT_AND,
+        BIT_OR,
+        BIT_XOR,
     };
 
     class EltwiseOpBase : public CUDABackendNode {
@@ -89,6 +94,10 @@ namespace cv { namespace dnn { namespace cuda4dnn {
                 case EltwiseOpType::MOD: kernels::eltwise_mod_2<T>(stream, output, input_x, input_y); break;
                 case EltwiseOpType::FMOD: kernels::eltwise_fmod_2<T>(stream, output, input_x, input_y); break;
                 case EltwiseOpType::POW: kernels::eltwise_pow_2<T>(stream, output, input_x, input_y); break;
+                case EltwiseOpType::PRELU: kernels::eltwise_prelu_2<T>(stream, output, input_x, input_y); break;
+                case EltwiseOpType::BIT_AND:
+                case EltwiseOpType::BIT_OR:
+                case EltwiseOpType::BIT_XOR: bitwise(output, input_x, input_y); break;
                 }
             } else if (inputs.size() == 1) {
                 auto input_0 = csl::viewOf<T>(inputs[0]);
@@ -123,10 +132,31 @@ namespace cv { namespace dnn { namespace cuda4dnn {
                     case EltwiseOpType::MOD: kernels::eltwise_mod_2<T>(stream, output, output, input); break;
                     case EltwiseOpType::FMOD: kernels::eltwise_fmod_2<T>(stream, output, output, input); break;
                     case EltwiseOpType::POW: kernels::eltwise_pow_2<T>(stream, output, output, input); break;
+                    case EltwiseOpType::PRELU: CV_Error(Error::StsBadArg, "PRelu takes exactly two inputs");
+                    case EltwiseOpType::BIT_AND:
+                    case EltwiseOpType::BIT_OR:
+                    case EltwiseOpType::BIT_XOR: bitwise(output, output, input); break;
                     }
                 }
             }
         }
+
+    private:
+        // the bitwise kernels exist for integer element types only
+        void bitwise(csl::TensorSpan<T> output, csl::TensorView<T> x, csl::TensorView<T> y)
+        {
+            if constexpr (std::is_integral<T>::value && !std::is_same<T, bool>::value) {
+                if (op == EltwiseOpType::BIT_AND)
+                    kernels::eltwise_bitand_2<T>(stream, output, x, y);
+                else if (op == EltwiseOpType::BIT_OR)
+                    kernels::eltwise_bitor_2<T>(stream, output, x, y);
+                else
+                    kernels::eltwise_bitxor_2<T>(stream, output, x, y);
+            } else {
+                CV_UNUSED(output); CV_UNUSED(x); CV_UNUSED(y);
+                CV_Error(Error::StsNotImplemented, "bitwise ops need integer tensors");
+            }
+        }
     };
 
 }}} /* namespace cv::dnn::cuda4dnn */

modules/dnn/src/cuda4dnn/primitives/gather_elements.hpp

@@ -0,0 +1,71 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+// Copyright (C) 2026, BigVision LLC, all rights reserved.
+// Third party copyrights are property of their respective owners.
+#ifndef OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_GATHER_ELEMENTS_HPP
+#define OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_GATHER_ELEMENTS_HPP
+
+#include "../../op_cuda.hpp"
+
+#include "../csl/stream.hpp"
+
+#include "../kernels/gather_elements.hpp"
+
+#include <opencv2/core.hpp>
+
+#include <cstddef>
+#include <cstdint>
+#include <utility>
+
+namespace cv { namespace dnn { namespace cuda4dnn {
+
+    template <class T>
+    class GatherElementsOp final : public CUDABackendNode {
+    public:
+        GatherElementsOp(csl::Stream stream_, int axis_)
+            : stream(std::move(stream_)), axis(axis_)
+        {
+        }
+
+        void forward(
+            const std::vector<UMat>& inputs,
+            const std::vector<UMat>& outputs,
+            csl::Workspace& workspace) override
+        {
+            CV_UNUSED(workspace);
+            CV_Assert(inputs.size() == 2 && outputs.size() == 1);
+
+            auto data = csl::viewOf<T>(inputs[0]);
+            auto output = csl::spanOf<T>(outputs[0]);
+
+            MatShape dataShape = cv::dnn::shape(inputs[0]);
+            MatShape indicesShape = cv::dnn::shape(inputs[1]);
+            const int dataDims = dataShape.dims;
+            const int axis_ = normalize_axis(axis, dataDims);
+
+            std::int64_t inner_size = 1;
+            for (int i = axis_ + 1; i < dataDims; i++)
+                inner_size *= dataShape[i];
+
+            const std::int64_t axis_size_data = dataShape[axis_];
+            const std::int64_t axis_size_indices = indicesShape[axis_];
+
+            csl::device::fast_divmod axis_inner_size(static_cast<int>(axis_size_indices * inner_size));
+            csl::device::fast_divmod inner_size_divmod(static_cast<int>(inner_size));
+
+            const std::size_t index_element_size = inputs[1].elemSize();
+            const void* indices = reinterpret_cast<const uchar*>(inputs[1].u->handle) + inputs[1].offset;
+
+            kernels::gather_elements<T>(stream, output, data, indices, index_element_size,
+                                        axis_size_data, inner_size, axis_inner_size, inner_size_divmod);
+        }
+
+    private:
+        csl::Stream stream;
+        int axis;
+    };
+
+}}} /* namespace cv::dnn::cuda4dnn */
+
+#endif /* OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_GATHER_ELEMENTS_HPP */

modules/dnn/src/cuda4dnn/primitives/gather_nd.hpp

@@ -0,0 +1,104 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+// Copyright (C) 2026, BigVision LLC, all rights reserved.
+// Third party copyrights are property of their respective owners.
+#ifndef OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_GATHER_ND_HPP
+#define OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_GATHER_ND_HPP
+
+#include "../../op_cuda.hpp"
+
+#include "../csl/stream.hpp"
+
+#include "../kernels/gather_nd.hpp"
+
+#include <opencv2/core.hpp>
+
+#include <cstddef>
+#include <cstdint>
+#include <utility>
+#include <vector>
+
+namespace cv { namespace dnn { namespace cuda4dnn {
+
+    template <class T>
+    class GatherNDOp final : public CUDABackendNode {
+    public:
+        GatherNDOp(csl::Stream stream_, int batch_dims_)
+            : stream(std::move(stream_)), batch_dims(batch_dims_)
+        {
+        }
+
+        void forward(
+            const std::vector<UMat>& inputs,
+            const std::vector<UMat>& outputs,
+            csl::Workspace& workspace) override
+        {
+            CV_UNUSED(workspace);
+            CV_Assert(inputs.size() == 2 && outputs.size() == 1);
+
+            auto data = csl::viewOf<T>(inputs[0]);
+            auto output = csl::spanOf<T>(outputs[0]);
+
+            const MatShape dataShape = cv::dnn::shape(inputs[0]);
+            const MatShape indicesShape = cv::dnn::shape(inputs[1]);
+            const int r = dataShape.dims;
+            const int q = indicesShape.dims;
+            CV_Assert(r >= 1 && q >= 1);
+
+            const std::int64_t last_index_dimension = indicesShape[q - 1];
+            CV_Assert(last_index_dimension >= 1 && batch_dims + last_index_dimension <= r);
+
+            std::vector<std::int64_t> data_strides(r);
+            data_strides[r - 1] = 1;
+            for (int i = r - 2; i >= 0; --i)
+                data_strides[i] = data_strides[i + 1] * dataShape[i + 1];
+
+            std::vector<std::int64_t> element_counts(last_index_dimension);
+            std::vector<std::int64_t> input_dims(last_index_dimension);
+            for (std::int64_t j = 0; j < last_index_dimension; ++j) {
+                element_counts[j] = data_strides[batch_dims + j];
+                input_dims[j] = dataShape[batch_dims + static_cast<int>(j)];
+            }
+
+            std::int64_t indices_total = 1;
+            for (int i = 0; i < q; i++)
+                indices_total *= indicesShape[i];
+            const std::int64_t outer_size = indices_total / last_index_dimension;
+
+            // an empty indices or output tensor is a legal no-op, not an error
+            if (outer_size <= 0 || output.size() == 0)
+                return;
+
+            const std::int64_t inner_size = static_cast<std::int64_t>(output.size()) / outer_size;
+
+            std::int64_t num_slices_per_batch = 1;
+            for (int d = batch_dims; d < q - 1; ++d)
+                num_slices_per_batch *= indicesShape[d];
+            if (num_slices_per_batch <= 0)
+                num_slices_per_batch = 1;
+
+            const std::int64_t batch_stride = batch_dims > 0 ? data_strides[batch_dims - 1] : 0;
+
+            const int idepth = inputs[1].depth();
+            if (idepth == CV_32S) {
+                kernels::gather_nd<T, std::int32_t>(stream, output, csl::viewOf<std::int32_t>(inputs[1]),
+                    data, last_index_dimension, element_counts, input_dims,
+                    inner_size, batch_stride, num_slices_per_batch);
+            } else if (idepth == CV_64S) {
+                kernels::gather_nd<T, std::int64_t>(stream, output, csl::viewOf<std::int64_t>(inputs[1]),
+                    data, last_index_dimension, element_counts, input_dims,
+                    inner_size, batch_stride, num_slices_per_batch);
+            } else {
+                CV_Error(Error::StsNotImplemented, "GatherND: indices must be CV_32S or CV_64S");
+            }
+        }
+
+    private:
+        csl::Stream stream;
+        int batch_dims;
+    };
+
+}}} /* namespace cv::dnn::cuda4dnn */
+
+#endif /* OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_GATHER_ND_HPP */

modules/dnn/src/cuda4dnn/primitives/grid_sample.hpp

@@ -0,0 +1,61 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+// Copyright (C) 2026, BigVision LLC, all rights reserved.
+// Third party copyrights are property of their respective owners.
+#ifndef OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_GRID_SAMPLE_HPP
+#define OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_GRID_SAMPLE_HPP
+
+#include "../../op_cuda.hpp"
+
+#include "../csl/stream.hpp"
+
+#include "../kernels/grid_sample.hpp"
+
+#include <opencv2/core.hpp>
+
+#include <utility>
+#include <vector>
+
+namespace cv { namespace dnn { namespace cuda4dnn {
+
+    template <class T>
+    class GridSampleOp final : public CUDABackendNode {
+    public:
+        GridSampleOp(csl::Stream stream_, const kernels::GridSampleParams& params_)
+            : stream(std::move(stream_)), params(params_) { }
+
+        void forward(
+            const std::vector<UMat>& inputs,
+            const std::vector<UMat>& outputs,
+            csl::Workspace& workspace) override
+        {
+            CV_UNUSED(workspace);
+            CV_Assert(inputs.size() == 2 && outputs.size() == 1);
+
+            auto input = csl::viewOf<T>(inputs[0]);
+            auto output = csl::spanOf<T>(outputs[0]);
+            auto grid = csl::viewOf<float>(inputs[1]);
+
+            MatShape inShape = cv::dnn::shape(inputs[0]);
+            MatShape outShape = cv::dnn::shape(outputs[0]);
+            CV_Assert(inShape.dims == 4 && outShape.dims == 4);
+
+            kernels::GridSampleParams p = params;
+            p.channels      = inShape[1];
+            p.height        = inShape[2];
+            p.width         = inShape[3];
+            p.output_height = outShape[2];
+            p.output_width  = outShape[3];
+
+            kernels::grid_sample<T>(stream, output, input, grid, p);
+        }
+
+    private:
+        csl::Stream stream;
+        kernels::GridSampleParams params;
+    };
+
+}}} /* namespace cv::dnn::cuda4dnn */
+
+#endif /* OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_GRID_SAMPLE_HPP */

modules/dnn/src/cuda4dnn/primitives/group_norm.hpp

@@ -28,27 +28,27 @@ namespace cv { namespace dnn { namespace cuda4dnn {
     public:
         using wrapper_type = GetCUDABackendWrapperType<T>;
 
-        GroupNormOp(csl::Stream stream_, float epsilon_, size_t loops, size_t num_groups)
+        GroupNormOp(csl::Stream stream_, float epsilon_, size_t loops, size_t num_groups, size_t group_elems)
             : stream(std::move(stream_)), epsilon(epsilon_), num_groups(num_groups) {
+            reduce_scratch = kernels::reduce_mean_sqr_sum_scratch(loops, group_elems);
             csl::WorkspaceBuilder builder;
             builder.require<float>(loops * num_groups); // mean and stdev for each group
             builder.require<float>(loops * num_groups);
+            if (reduce_scratch)
+                builder.require<float>(reduce_scratch);
             scratch_mem_in_bytes = builder.required_workspace_size();
         }
 
-        void forward(const std::vector<cv::Ptr<BackendWrapper>>& inputs,
-                     const std::vector<cv::Ptr<BackendWrapper>>& outputs,
+        void forward(const std::vector<UMat>& inputs,
+                     const std::vector<UMat>& outputs,
                      csl::Workspace& workspace) override {
-            auto input_wrapper = inputs[0].dynamicCast<wrapper_type>();
-            auto scale_wrapper = inputs[1].dynamicCast<wrapper_type>();
-            auto bias_wrapper = inputs[2].dynamicCast<wrapper_type>();
+            CV_Assert(inputs.size() == 3 && outputs.size() == 1);
 
-            auto input = input_wrapper->getView();
-            auto scale = scale_wrapper->getView();
-            auto bias = bias_wrapper->getView();
+            auto input = csl::viewOf<T>(inputs[0]);
+            auto scale = csl::viewOf<T>(inputs[1]);
+            auto bias = csl::viewOf<T>(inputs[2]);
 
-            auto output_wrapper = outputs[0].dynamicCast<wrapper_type>();
-            auto output = output_wrapper->getSpan();
+            auto output = csl::spanOf<T>(outputs[0]);
 
             auto C = input.get_axis_size(1);
             auto loops = input.size_range(0, 2);
@@ -64,8 +64,9 @@ namespace cv { namespace dnn { namespace cuda4dnn {
                 // reduce_mean_sqr_sum writes every row, so no zero-fill is needed
                 auto mean = ws_allocator.get_span<float>(loops / group_size);
                 auto stdev = ws_allocator.get_span<float>(loops / group_size);
+                auto partials = reduce_scratch ? ws_allocator.get_span<float>(reduce_scratch) : csl::Span<float>();
 
-                kernels::reduce_mean_sqr_sum<T>(stream, mean, stdev, input, norm_size * group_size);
+                kernels::reduce_mean_sqr_sum<T>(stream, mean, stdev, input, norm_size * group_size, partials);
                 kernels::compute_normalization_scale(stream, stdev, mean, stdev, norm_size * group_size, epsilon);
                 kernels::normalize_mean_variance_groupwise<T>(stream, output, input, scale, bias, mean, stdev, norm_size, C, num_groups, group_size);
             }
@@ -77,6 +78,7 @@ namespace cv { namespace dnn { namespace cuda4dnn {
         csl::Stream stream;
         float epsilon;
         std::size_t num_groups;
+        std::size_t reduce_scratch;
         std::size_t scratch_mem_in_bytes;
     };
 

modules/dnn/src/cuda4dnn/primitives/instance_norm.hpp

@@ -28,27 +28,27 @@ namespace cv { namespace dnn { namespace cuda4dnn {
      public:
         using wrapper_type = GetCUDABackendWrapperType<T>;
 
-        InstanceNormOp(csl::Stream stream_, float epsilon_, size_t loops)
+        InstanceNormOp(csl::Stream stream_, float epsilon_, size_t loops, size_t norm_size)
             : stream(std::move(stream_)), epsilon(epsilon_) {
+            reduce_scratch = kernels::reduce_mean_sqr_sum_scratch(loops, norm_size);
             csl::WorkspaceBuilder builder;
             builder.require<float>(loops);
             builder.require<float>(loops);
+            if (reduce_scratch)
+                builder.require<float>(reduce_scratch);
             scratch_mem_in_bytes = builder.required_workspace_size();
         }
 
-        void forward(const std::vector<cv::Ptr<BackendWrapper>>& inputs,
-                     const std::vector<cv::Ptr<BackendWrapper>>& outputs,
+        void forward(const std::vector<UMat>& inputs,
+                     const std::vector<UMat>& outputs,
                      csl::Workspace& workspace) override {
-            auto input_wrapper = inputs[0].dynamicCast<wrapper_type>();
-            auto scale_wrapper = inputs[1].dynamicCast<wrapper_type>();
-            auto bias_wrapper = inputs[2].dynamicCast<wrapper_type>();
+            CV_Assert(inputs.size() == 3 && outputs.size() == 1);
 
-            auto input = input_wrapper->getView();
-            auto scale = scale_wrapper->getView();
-            auto bias = bias_wrapper->getView();
+            auto input = csl::viewOf<T>(inputs[0]);
+            auto scale = csl::viewOf<T>(inputs[1]);
+            auto bias = csl::viewOf<T>(inputs[2]);
 
-            auto output_wrapper = outputs[0].dynamicCast<wrapper_type>();
-            auto output = output_wrapper->getSpan();
+            auto output = csl::spanOf<T>(outputs[0]);
 
             auto C = input.get_axis_size(1);
             auto loops = input.size_range(0, 2);
@@ -62,8 +62,9 @@ namespace cv { namespace dnn { namespace cuda4dnn {
                 // reduce_mean_sqr_sum writes every row, so no zero-fill is needed
                 auto mean = ws_allocator.get_span<float>(loops);
                 auto stdev = ws_allocator.get_span<float>(loops);
+                auto partials = reduce_scratch ? ws_allocator.get_span<float>(reduce_scratch) : csl::Span<float>();
 
-                kernels::reduce_mean_sqr_sum<T>(stream, mean, stdev, input, norm_size);
+                kernels::reduce_mean_sqr_sum<T>(stream, mean, stdev, input, norm_size, partials);
                 kernels::compute_normalization_scale(stream, stdev, mean, stdev, norm_size, epsilon);
                 kernels::normalize_mean_variance_channelwise<T>(stream, output, input, scale, bias, mean, stdev, norm_size, C);
             }
@@ -76,6 +77,7 @@ namespace cv { namespace dnn { namespace cuda4dnn {
 
         float epsilon;
 
+        std::size_t reduce_scratch;
         std::size_t scratch_mem_in_bytes;
     };
 

modules/dnn/src/cuda4dnn/primitives/layer_norm.hpp

@@ -77,7 +77,8 @@ namespace cv { namespace dnn { namespace cuda4dnn {
 
             auto loops = input.size_range(0, axis);
             auto norm_size = input.size_range(axis, input.rank());
-            if (norm_size == 1) {
+            // with a size-one axis the result is 0, or the bias when there is one
+            if (norm_size == 1 && bias.empty()) {
                 kernels::fill<T>(stream, output, 0.f);
                 return;
             } else {
@@ -114,7 +115,9 @@ namespace cv { namespace dnn { namespace cuda4dnn {
 
             auto loops = input.size_range(0, axis);
             auto norm_size = input.size_range(axis, input.rank());
-            if (norm_size == 1) {
+            // with a size-one axis the result is 0, or the bias when there is one;
+            // the optional Mean/InvStdDev outputs still need the statistics
+            if (norm_size == 1 && outputs.size() == 1 && bias.empty()) {
                 kernels::fill<T>(stream, output, 0.f);
                 return;
             }
@@ -131,6 +134,12 @@ namespace cv { namespace dnn { namespace cuda4dnn {
                 kernels::normalize_mean_variance_layernorm<T>(stream, output, input, scale, bias, mean, inv_stddev, norm_size);
             else
                 kernels::normalize_mean_variance_layernorm<T>(stream, output, input, scale, mean, inv_stddev, norm_size);
+
+            // ONNX Mean and InvStdDev are float per row, which is what the scratch holds
+            if (outputs.size() >= 2 && !outputs[1].empty())
+                csl::memcpy<float>(csl::spanOf<float>(outputs[1]).get(), mean.data(), loops, stream);
+            if (outputs.size() >= 3 && !outputs[2].empty())
+                csl::memcpy<float>(csl::spanOf<float>(outputs[2]).get(), inv_stddev.data(), loops, stream);
         }
 
         std::size_t get_workspace_memory_in_bytes() const noexcept override { return scratch_mem_in_bytes; }

modules/dnn/src/cuda4dnn/primitives/lrn.hpp

@@ -10,8 +10,11 @@
 #include "../csl/cudnn.hpp"
 #include "../csl/tensor_ops.hpp"
 
+#include "../kernels/normalize.hpp"
+
 #include <cstddef>
 #include <utility>
+#include <vector>
 
 namespace cv { namespace dnn { namespace cuda4dnn {
 
@@ -70,6 +73,38 @@ namespace cv { namespace dnn { namespace cuda4dnn {
         std::size_t scratch_mem_in_bytes;
     };
 
+    // across-channel LRN without legacy cuDNN; matches the CPU ChannelLRN
+    template <class T>
+    class LRNAcrossChannelsOp final : public CUDABackendNode {
+    public:
+        LRNAcrossChannelsOp(csl::Stream stream_, int local_size, float alpha_scaled_, float beta_, float bias_)
+            : stream(std::move(stream_)), radius((local_size - 1) / 2), alpha_scaled(alpha_scaled_), beta(beta_), bias(bias_) { }
+
+        void forward(
+            const std::vector<UMat>& inputs,
+            const std::vector<UMat>& outputs,
+            csl::Workspace& workspace) override
+        {
+            CV_UNUSED(workspace);
+            CV_Assert(inputs.size() == outputs.size());
+            for (size_t i = 0; i < inputs.size(); i++) {
+                auto input = csl::viewOf<T>(inputs[i]);
+                auto output = csl::spanOf<T>(outputs[i]);
+                CV_Assert(input.rank() >= 2);
+                const std::size_t batch = input.get_axis_size(0), channels = input.get_axis_size(1);
+                const std::size_t plane = input.size_range(2, input.rank());
+                kernels::lrn_across_channels<T>(stream, csl::Span<T>(output.get(), output.size()),
+                                                csl::View<T>(input.get(), input.size()),
+                                                batch, channels, plane, radius, alpha_scaled, bias, beta);
+            }
+        }
+
+    private:
+        csl::Stream stream;
+        int radius;
+        float alpha_scaled, beta, bias;
+    };
+
 }}} /* namespace cv::dnn::cuda4dnn */
 
 #endif /* OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_LRN_HPP */

modules/dnn/src/cuda4dnn/primitives/normalize_bbox.hpp

@@ -87,17 +87,14 @@ namespace cv { namespace dnn { namespace cuda4dnn {
         }
 
         void forward(
-            const std::vector<cv::Ptr<BackendWrapper>>& inputs,
-            const std::vector<cv::Ptr<BackendWrapper>>& outputs,
+            const std::vector<UMat>& inputs,
+            const std::vector<UMat>& outputs,
             csl::Workspace& workspace) override
         {
             CV_Assert(inputs.size() == 1 && outputs.size() == 1);
 
-            auto input_wrapper = inputs[0].dynamicCast<wrapper_type>();
-            auto input = input_wrapper->getView();
-
-            auto output_wrapper = outputs[0].dynamicCast<wrapper_type>();
-            auto output = output_wrapper->getSpan();
+            auto input = csl::viewOf<T>(inputs[0]);
+            auto output = csl::spanOf<T>(outputs[0]);
 
             std::size_t outer_size = input.size_range(0, axis_start);
             std::size_t mid_size = input.size_range(axis_start, axis_end);

modules/dnn/src/cuda4dnn/primitives/one_hot.hpp

@@ -0,0 +1,76 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+// Copyright (C) 2026, BigVision LLC, all rights reserved.
+// Third party copyrights are property of their respective owners.
+#ifndef OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_ONE_HOT_HPP
+#define OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_ONE_HOT_HPP
+
+#include "../../op_cuda.hpp"
+
+#include "../csl/stream.hpp"
+
+#include "../kernels/one_hot.hpp"
+
+#include <opencv2/core.hpp>
+
+#include <cstddef>
+#include <vector>
+#include <utility>
+
+namespace cv { namespace dnn { namespace cuda4dnn {
+
+    template <class T, class TIdx>
+    class OneHotOp final : public CUDABackendNode {
+    public:
+        OneHotOp(csl::Stream stream_, int axis_, double off_, double on_, bool runtimeValues_)
+            : stream(std::move(stream_)), axis(axis_),
+              offValue(static_cast<T>(off_)), onValue(static_cast<T>(on_)),
+              runtimeValues(runtimeValues_)
+        {
+        }
+
+        void forward(
+            const std::vector<UMat>& inputs,
+            const std::vector<UMat>& outputs,
+            csl::Workspace& workspace) override
+        {
+            CV_UNUSED(workspace);
+            CV_Assert(inputs.size() == 3 && outputs.size() == 1);
+
+            MatShape outShape = cv::dnn::shape(outputs[0]);
+            const int rank = outShape.dims;
+            const int a = axis < 0 ? axis + rank : axis;
+            CV_Assert(0 <= a && a < rank);
+            if (outputs[0].total() == 0)
+                return;
+
+            if (runtimeValues && !inputs[2].empty()) {
+                Mat values, values64;
+                inputs[2].copyTo(values);
+                CV_Assert(values.total() == 2);
+                values.convertTo(values64, CV_64F);
+                offValue = static_cast<T>(values64.ptr<double>()[0]);
+                onValue = static_cast<T>(values64.ptr<double>()[1]);
+            }
+
+            std::size_t inner = 1;
+            for (int i = a + 1; i < rank; i++)
+                inner *= static_cast<std::size_t>(outShape[i]);
+
+            auto indices = csl::viewOf<TIdx>(inputs[0]);
+            auto output = csl::spanOf<T>(outputs[0]);
+            kernels::one_hot<T, TIdx>(stream, output, indices, static_cast<std::size_t>(outShape[a]), inner,
+                                      offValue, onValue);
+        }
+
+    private:
+        csl::Stream stream;
+        int axis;
+        T offValue, onValue;
+        bool runtimeValues;
+    };
+
+}}} /* namespace cv::dnn::cuda4dnn */
+
+#endif /* OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_ONE_HOT_HPP */

modules/dnn/src/cuda4dnn/primitives/padding.hpp

@@ -17,6 +17,7 @@
 #include <opencv2/core.hpp>
 
 #include <cstddef>
+#include <cstring>
 #include <vector>
 #include <algorithm>
 #include <utility>
@@ -113,6 +114,37 @@ namespace cv { namespace dnn { namespace cuda4dnn {
         std::vector<cv::Range> dstRanges;
     };
 
+    // ONNX Pad: any rank, constant/reflect/edge/wrap, negative pads crop
+    template <class T>
+    class Pad2Op final : public CUDABackendNode {
+    public:
+        // value_bits holds the pad value already converted to the tensor's element type
+        Pad2Op(csl::Stream stream_, std::vector<int> pads_begin_, kernels::PadMode mode_, const void* value_bits)
+            : stream(std::move(stream_)), pads_begin(std::move(pads_begin_)), mode(mode_)
+        {
+            std::memcpy(&value, value_bits, sizeof(T));
+        }
+
+        void forward(
+            const std::vector<UMat>& inputs,
+            const std::vector<UMat>& outputs,
+            csl::Workspace& workspace) override
+        {
+            CV_UNUSED(workspace);
+            CV_Assert(!inputs.empty() && outputs.size() == 1);
+
+            auto input = csl::viewOf<T>(inputs[0]);
+            auto output = csl::spanOf<T>(outputs[0]);
+            kernels::pad_nd<T>(stream, output, input, pads_begin, mode, value);
+        }
+
+    private:
+        csl::Stream stream;
+        std::vector<int> pads_begin;
+        kernels::PadMode mode;
+        T value;
+    };
+
 }}} /* namespace cv::dnn::cuda4dnn */
 
 #endif /* OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_PADDING_HPP */

modules/dnn/src/cuda4dnn/primitives/pooling.hpp

@@ -17,9 +17,8 @@
 #include "../csl/tensor.hpp"
 #include "../csl/tensor_ops.hpp"
 
-#if defined(HAVE_CUDNNJIT) && !defined(HAVE_CUDNN)
+// max pooling with left padding uses the hand kernel in every build
 #include "../kernels/max_unpooling.hpp"
-#endif
 
 #include <opencv2/core.hpp>
 
@@ -181,13 +180,21 @@ namespace cv { namespace dnn { namespace cuda4dnn {
                 if (std::any_of(std::begin(padding_left), std::end(padding_left), is_not_zero))
                 {
 #if !defined(HAVE_CUDNNJIT) || defined(HAVE_CUDNN)
-                    /* there is padding on the left and we are forced to transform */
-                    auto transformed_input_shape = input_shape;
-                    for (int i = 0; i < rank; i++)
-                        transformed_input_shape[i] += padding_left[i] + padding_right[i];
+                    // the transform pads with zeros, which wins a max over negative borders; the index kernel skips them
+                    if (config.poolMode == PoolingConfiguration::PoolingMode::MAX)
+                    {
+                        useIndexKernel = true;
+                    }
+                    else
+                    {
+                        /* there is padding on the left and we are forced to transform */
+                        auto transformed_input_shape = input_shape;
+                        for (int i = 0; i < rank; i++)
+                            transformed_input_shape[i] += padding_left[i] + padding_right[i];
 
-                    transformedInput.resize(std::begin(transformed_input_shape), std::end(transformed_input_shape));
-                    inputTransformer = csl::TensorTransform<T>(cudnnHandle, padding_left, padding_right);
+                        transformedInput.resize(std::begin(transformed_input_shape), std::end(transformed_input_shape));
+                        inputTransformer = csl::TensorTransform<T>(cudnnHandle, padding_left, padding_right);
+                    }
 #endif
                 }
             }
@@ -237,12 +244,20 @@ namespace cv { namespace dnn { namespace cuda4dnn {
 #if defined(HAVE_CUDNNJIT) && !defined(HAVE_CUDNN)
             // JIT builds route average pooling to AveragePoolingOp, so only MAX reaches here
             CV_Assert(config.poolMode == PoolingConfiguration::PoolingMode::MAX);
-            poolWindow.assign(std::begin(window_size), std::end(window_size));
-            poolStrides.assign(std::begin(strides), std::end(strides));
-            for (std::size_t i = 0; i < window_size.size(); i++)
-                poolPadsBegin.push_back(common_padding[i + 2] + padding_left[i + 2]);
-#else
-            pooler = csl::Pooling<T>(cudnnHandle, params);
+            useIndexKernel = true;
+#endif
+            if (useIndexKernel)
+            {
+                poolWindow.assign(std::begin(window_size), std::end(window_size));
+                poolStrides.assign(std::begin(strides), std::end(strides));
+                for (std::size_t i = 0; i < window_size.size(); i++)
+                    poolPadsBegin.push_back(common_padding[i + 2] + padding_left[i + 2]);
+            }
+#if !defined(HAVE_CUDNNJIT) || defined(HAVE_CUDNN)
+            else
+            {
+                pooler = csl::Pooling<T>(cudnnHandle, params);
+            }
 #endif
         }
 
@@ -256,11 +271,14 @@ namespace cv { namespace dnn { namespace cuda4dnn {
             auto input = csl::viewOf<T>(inputs[0]);
             auto output = csl::spanOf<T>(outputs[0]);
 
-#if defined(HAVE_CUDNNJIT) && !defined(HAVE_CUDNN)
             CV_UNUSED(workspace);
-            kernels::max_pooling_with_indices<T, std::int64_t>(stream, output, csl::TensorSpan<std::int64_t>(), input,
-                                                               poolWindow, poolStrides, poolPadsBegin);
-#else
+            if (useIndexKernel)
+            {
+                kernels::max_pooling_with_indices<T, std::int64_t>(stream, output, csl::TensorSpan<std::int64_t>(), input,
+                                                                   poolWindow, poolStrides, poolPadsBegin);
+                return;
+            }
+#if !defined(HAVE_CUDNNJIT) || defined(HAVE_CUDNN)
             if (!transformedInput.empty())
             {
                 inputTransformer.transform(input, transformedInput);
@@ -275,9 +293,10 @@ namespace cv { namespace dnn { namespace cuda4dnn {
         csl::Stream stream;
         csl::cudnn::Handle cudnnHandle;
 
-#if defined(HAVE_CUDNNJIT) && !defined(HAVE_CUDNN)
+        // max pooling through the hand kernel: always in JIT builds, for left padding in full cuDNN builds
+        bool useIndexKernel = false;
         std::vector<std::size_t> poolWindow, poolStrides, poolPadsBegin;
-#else
+#if !defined(HAVE_CUDNNJIT) || defined(HAVE_CUDNN)
         csl::Pooling<T> pooler;
 
         csl::Tensor<T> transformedInput;

modules/dnn/src/cuda4dnn/primitives/quantize_dequantize.hpp

@@ -0,0 +1,138 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+
+#ifndef OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_QUANTIZE_DEQUANTIZE_HPP
+#define OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_QUANTIZE_DEQUANTIZE_HPP
+
+#include "../../op_cuda.hpp"
+
+#include "../csl/stream.hpp"
+#include "../csl/workspace.hpp"
+#include "../kernels/quantize_dequantize.hpp"
+
+#include <opencv2/dnn/shape_utils.hpp>
+
+#include <cstdint>
+#include <vector>
+
+namespace cv { namespace dnn { namespace cuda4dnn {
+
+    // Per-tensor and per-axis only (block_size == 0); gated by the layer's supportBackend.
+    template <class T>
+    class DequantizeLinearOp final : public CUDABackendNode {
+    public:
+        DequantizeLinearOp(csl::Stream stream_, int axis_) : stream(std::move(stream_)), axis(axis_) { }
+
+        void forward(
+            const std::vector<UMat>& inputs,
+            const std::vector<UMat>& outputs,
+            csl::Workspace& workspace) override
+        {
+            CV_UNUSED(workspace);
+            CV_Assert(inputs.size() == 2 || inputs.size() == 3);
+            CV_Assert(outputs.size() == 1);
+
+            auto input = csl::viewOf<T>(inputs[0]);
+            auto scale = csl::viewOf<float>(inputs[1]);
+            auto output = csl::spanOf<float>(outputs[0]);
+            csl::View<T> zero_point = inputs.size() == 3 ? csl::viewOf<T>(inputs[2]) : csl::View<T>();
+            CV_Assert(zero_point.empty() || zero_point.size() == scale.size());
+
+            MatShape inpshape = shape(inputs[0]);
+            int ndims = inpshape.dims;
+            int ax = ndims > 0 ? normalize_axis(axis, ndims) : 0;
+            std::size_t sz_a = scale.size() > 1 ? (std::size_t)inpshape[ax] : 1;
+            CV_Assert(scale.size() == 1 || scale.size() == sz_a);
+            std::size_t slice_size = 1;
+            if (sz_a > 1)
+                for (int i = ax + 1; i < ndims; i++)
+                    slice_size *= inpshape[i];
+
+            kernels::dequantize_linear<T>(stream, output, input, scale, zero_point, sz_a, slice_size);
+        }
+
+    private:
+        csl::Stream stream;
+        int axis;
+    };
+
+    template <class T>
+    class QuantizeLinearOp final : public CUDABackendNode {
+    public:
+        QuantizeLinearOp(csl::Stream stream_, int axis_) : stream(std::move(stream_)), axis(axis_) { }
+
+        void forward(
+            const std::vector<UMat>& inputs,
+            const std::vector<UMat>& outputs,
+            csl::Workspace& workspace) override
+        {
+            CV_UNUSED(workspace);
+            CV_Assert(inputs.size() == 2 || inputs.size() == 3);
+            CV_Assert(outputs.size() == 1);
+
+            auto input = csl::viewOf<float>(inputs[0]);
+            auto scale = csl::viewOf<float>(inputs[1]);
+            auto output = csl::spanOf<T>(outputs[0]);
+            csl::View<T> zero_point = inputs.size() == 3 ? csl::viewOf<T>(inputs[2]) : csl::View<T>();
+            CV_Assert(zero_point.empty() || zero_point.size() == scale.size());
+
+            MatShape inpshape = shape(inputs[0]);
+            int ndims = inpshape.dims;
+            int ax = ndims > 0 ? normalize_axis(axis, ndims) : 0;
+            std::size_t sz_a = scale.size() > 1 ? (std::size_t)inpshape[ax] : 1;
+            CV_Assert(scale.size() == 1 || scale.size() == sz_a);
+            std::size_t slice_size = 1;
+            if (sz_a > 1)
+                for (int i = ax + 1; i < ndims; i++)
+                    slice_size *= inpshape[i];
+
+            kernels::quantize_linear<T>(stream, output, input, scale, zero_point, sz_a, slice_size);
+        }
+
+    private:
+        csl::Stream stream;
+        int axis;
+    };
+
+    // ONNX DynamicQuantizeLinear: float32 x -> uint8 y, scalar y_scale and y_zero_point, all on the device.
+    class DynamicQuantizeLinearOp final : public CUDABackendNode {
+    public:
+        DynamicQuantizeLinearOp(csl::Stream stream_, std::size_t total) : stream(std::move(stream_))
+        {
+            partials = kernels::dynamic_quantize_partials(total);
+            csl::WorkspaceBuilder builder;
+            builder.require<float>(2 * partials);
+            scratch_mem_in_bytes = builder.required_workspace_size();
+        }
+
+        void forward(
+            const std::vector<UMat>& inputs,
+            const std::vector<UMat>& outputs,
+            csl::Workspace& workspace) override
+        {
+            CV_Assert(inputs.size() == 1 && outputs.size() == 3);
+
+            auto input = csl::viewOf<float>(inputs[0]);
+            auto output = csl::spanOf<std::uint8_t>(outputs[0]);
+            auto scale = csl::spanOf<float>(outputs[1]);
+            auto zero_point = csl::spanOf<std::uint8_t>(outputs[2]);
+
+            auto allocator = csl::WorkspaceAllocator(workspace);
+            auto scratch = allocator.get_span<float>(2 * partials);
+            kernels::dynamic_quantize_scale(stream, scale, zero_point, input, scratch);
+            // same per-tensor formula as QuantizeLinear, with the scale and zero point read on the device
+            kernels::quantize_linear<std::uint8_t>(stream, output, input, scale, zero_point, 1, 1);
+        }
+
+        std::size_t get_workspace_memory_in_bytes() const noexcept override { return scratch_mem_in_bytes; }
+
+    private:
+        csl::Stream stream;
+        std::size_t partials;
+        std::size_t scratch_mem_in_bytes;
+    };
+
+}}} /* namespace cv::dnn::cuda4dnn */
+
+#endif /* OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_QUANTIZE_DEQUANTIZE_HPP */

modules/dnn/src/cuda4dnn/primitives/range.hpp

@@ -0,0 +1,52 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+// Copyright (C) 2026, BigVision LLC, all rights reserved.
+// Third party copyrights are property of their respective owners.
+#ifndef OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_RANGE_HPP
+#define OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_RANGE_HPP
+
+#include "../../op_cuda.hpp"
+
+#include "../csl/stream.hpp"
+
+#include "../kernels/range.hpp"
+
+#include <opencv2/core.hpp>
+
+#include <utility>
+
+namespace cv { namespace dnn { namespace cuda4dnn {
+
+    template <class T>
+    class RangeOp final : public CUDABackendNode {
+    public:
+        RangeOp(csl::Stream stream_) : stream(std::move(stream_)) { }
+
+        void forward(
+            const std::vector<UMat>& inputs,
+            const std::vector<UMat>& outputs,
+            csl::Workspace& workspace) override
+        {
+            CV_UNUSED(workspace);
+            CV_Assert(inputs.size() == 3 && outputs.size() == 1);
+            CV_Assert(inputs[0].total() == 1 && inputs[2].total() == 1);
+            CV_Assert(inputs[0].u && inputs[2].u);
+
+            // start/delta are dereferenced inside the kernel: no host round trip, no stall
+            auto devPtr = [](const UMat& m) {
+                return reinterpret_cast<const T*>(
+                    reinterpret_cast<const uchar*>(m.u->handle) + m.offset);
+            };
+
+            auto output = csl::spanOf<T>(outputs[0]);
+            kernels::range<T>(stream, output, devPtr(inputs[0]), devPtr(inputs[2]));
+        }
+
+    private:
+        csl::Stream stream;
+    };
+
+}}} /* namespace cv::dnn::cuda4dnn */
+
+#endif /* OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_RANGE_HPP */

modules/dnn/src/cuda4dnn/primitives/resize.hpp

@@ -10,8 +10,10 @@
 #include "../csl/stream.hpp"
 
 #include "../kernels/resize.hpp"
+#include "../csl/span.hpp"
 
 #include <utility>
+#include <vector>
 
 namespace cv { namespace dnn { namespace cuda4dnn {
 
@@ -40,18 +42,15 @@ namespace cv { namespace dnn { namespace cuda4dnn {
         }
 
         void forward(
-            const std::vector<cv::Ptr<BackendWrapper>>& inputs,
-            const std::vector<cv::Ptr<BackendWrapper>>& outputs,
+            const std::vector<UMat>& inputs,
+            const std::vector<UMat>& outputs,
             csl::Workspace& workspace) override
         {
-            // sometimes the target shape is taken from the second input; we don't use it however
-            CV_Assert((inputs.size() == 1 || inputs.size() == 2) && outputs.size() == 1);
+            // roi/scales/sizes may follow the data input; the output shape already carries them
+            CV_Assert(inputs.size() >= 1 && outputs.size() == 1);
 
-            auto input_wrapper = inputs[0].dynamicCast<wrapper_type>();
-            auto input = input_wrapper->getView();
-
-            auto output_wrapper = outputs[0].dynamicCast<wrapper_type>();
-            auto output = output_wrapper->getSpan();
+            auto input = csl::viewOf<T>(inputs[0]);
+            auto output = csl::spanOf<T>(outputs[0]);
 
             const auto compute_scale = [this](std::size_t input_size, std::size_t output_size) {
                 return (align_corners && output_size > 1) ?
@@ -76,6 +75,63 @@ namespace cv { namespace dnn { namespace cuda4dnn {
         bool align_corners, half_pixel_centers;
     };
 
+    // per-axis taps built on the host by the CPU Resize2 builders; idx -1 skips a tap
+    struct ResizeTables {
+        int taps_y = 1, taps_x = 1;
+        std::vector<int> idx_y, idx_x, oob_y, oob_x;
+        std::vector<float> w_y, w_x;
+        float extrapolation_value = 0.f;
+    };
+
+    // every ONNX Resize mode as one gather; the tables carry the coordinate and filter math
+    template <class T>
+    class ResizeTablesOp final : public CUDABackendNode {
+    public:
+        ResizeTablesOp(csl::Stream stream_, const ResizeTables& t)
+            : stream(std::move(stream_)), taps_y(t.taps_y), taps_x(t.taps_x), extrapolation(t.extrapolation_value)
+        {
+            idx_y = upload(t.idx_y); w_y = upload(t.w_y); oob_y = upload(t.oob_y);
+            idx_x = upload(t.idx_x); w_x = upload(t.w_x); oob_x = upload(t.oob_x);
+        }
+
+        void forward(
+            const std::vector<UMat>& inputs,
+            const std::vector<UMat>& outputs,
+            csl::Workspace& workspace) override
+        {
+            CV_UNUSED(workspace);
+            CV_Assert(inputs.size() >= 1 && outputs.size() == 1);
+            auto input = csl::viewOf<T>(inputs[0]);
+            auto output = csl::spanOf<T>(outputs[0]);
+            kernels::resize_separable<T>(stream, output, input,
+                                         view<int>(idx_y), view<float>(w_y), view<int>(oob_y), taps_y,
+                                         view<int>(idx_x), view<float>(w_x), view<int>(oob_x), taps_x, extrapolation);
+        }
+
+    private:
+        // device copy on the CUDA UMat allocator, like every other device tensor
+        template <class V>
+        static UMat upload(const std::vector<V>& v)
+        {
+            CV_Assert(!v.empty());
+            Mat host(1, (int)v.size(), DataType<V>::type, const_cast<V*>(v.data()));
+            UMat u;
+            u.allocator = cv::cuda::getCudaAllocator();
+            u.fitSameSize(host, host.type());
+            checkCudaAllocated(u);
+            host.copyTo(u);
+            return u;
+        }
+
+        template <class V>
+        static csl::View<V> view(const UMat& u) { return csl::View<V>(csl::viewOf<V>(u).get(), u.total()); }
+
+        csl::Stream stream;
+        int taps_y, taps_x;
+        float extrapolation;
+        UMat idx_y, oob_y, idx_x, oob_x, w_y, w_x;
+    };
+
 }}} /* namespace cv::dnn::cuda4dnn */
 
 #endif /* OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_RESIZE_HPP */

modules/dnn/src/cuda4dnn/primitives/rms_norm.hpp

@@ -0,0 +1,61 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+// Copyright (C) 2026, BigVision LLC, all rights reserved.
+// Third party copyrights are property of their respective owners.
+#ifndef OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_RMS_NORM_HPP
+#define OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_RMS_NORM_HPP
+
+#include "../../op_cuda.hpp"
+
+#include "../csl/stream.hpp"
+
+#include "../kernels/rms_norm.hpp"
+
+#include <opencv2/core.hpp>
+#include <opencv2/dnn/shape_utils.hpp>
+
+#include <cstddef>
+#include <utility>
+#include <vector>
+
+namespace cv { namespace dnn { namespace cuda4dnn {
+
+    template <class T>
+    class RMSNormOp final : public CUDABackendNode {
+    public:
+        RMSNormOp(csl::Stream stream_, int axis_, float epsilon_)
+            : stream(std::move(stream_)), axis(axis_), epsilon(epsilon_) { }
+
+        void forward(
+            const std::vector<UMat>& inputs,
+            const std::vector<UMat>& outputs,
+            csl::Workspace& workspace) override
+        {
+            CV_UNUSED(workspace);
+            CV_Assert(inputs.size() >= 2 && outputs.size() == 1);
+
+            auto input = csl::viewOf<T>(inputs[0]);
+            auto scale = csl::viewOf<T>(inputs[1]);
+            auto output = csl::spanOf<T>(outputs[0]);
+
+            MatShape inShape = cv::dnn::shape(inputs[0]);
+            const int a = normalize_axis(axis, inShape.dims);
+            CV_Assert(a >= 0 && a < inShape.dims);
+
+            std::size_t loops = 1, norm_size = 1;
+            for (int i = 0; i < a; i++) loops *= static_cast<std::size_t>(inShape[i]);
+            for (int i = a; i < inShape.dims; i++) norm_size *= static_cast<std::size_t>(inShape[i]);
+
+            kernels::rms_norm<T>(stream, output, input, scale, loops, norm_size, epsilon);
+        }
+
+    private:
+        csl::Stream stream;
+        int axis;
+        float epsilon;
+    };
+
+}}} /* namespace cv::dnn::cuda4dnn */
+
+#endif /* OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_RMS_NORM_HPP */

modules/dnn/src/cuda4dnn/primitives/roi_align.hpp

@@ -0,0 +1,68 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+// Copyright (C) 2026, BigVision LLC, all rights reserved.
+// Third party copyrights are property of their respective owners.
+#ifndef OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_ROI_ALIGN_HPP
+#define OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_ROI_ALIGN_HPP
+
+#include "../../op_cuda.hpp"
+
+#include "../csl/stream.hpp"
+
+#include "../kernels/roi_align.hpp"
+
+#include <opencv2/core.hpp>
+
+#include <cstddef>
+#include <utility>
+#include <vector>
+
+namespace cv { namespace dnn { namespace cuda4dnn {
+
+    template <class T>
+    class RoiAlignOp final : public CUDABackendNode {
+    public:
+        RoiAlignOp(csl::Stream stream_, const kernels::RoiAlignParams& params_)
+            : stream(std::move(stream_)), params(params_) { }
+
+        void forward(
+            const std::vector<UMat>& inputs,
+            const std::vector<UMat>& outputs,
+            csl::Workspace& workspace) override
+        {
+            CV_UNUSED(workspace);
+            CV_Assert(inputs.size() == 3 && outputs.size() == 1);
+
+            auto input = csl::viewOf<T>(inputs[0]);
+            auto output = csl::spanOf<T>(outputs[0]);
+            auto rois = csl::viewOf<float>(inputs[1]);
+
+            MatShape inShape = cv::dnn::shape(inputs[0]);
+            CV_Assert(inShape.dims == 4);
+
+            /* The spatial extents are only known once shapes are resolved, so the
+             * per-run fields are filled here and the attribute-derived ones come from
+             * the layer. */
+            kernels::RoiAlignParams p = params;
+            p.batch_size = inShape[0];
+            p.channels = inShape[1];
+            p.height   = inShape[2];
+            p.width    = inShape[3];
+
+            const std::size_t batch_index_element_size = inputs[2].elemSize();
+            // device pointer, same idiom the gather primitive uses for its index tensor
+            const void* batch_indices = reinterpret_cast<const uchar*>(inputs[2].u->handle) + inputs[2].offset;
+
+            kernels::roi_align<T>(stream, output, input, rois,
+                                  batch_indices, batch_index_element_size, p);
+        }
+
+    private:
+        csl::Stream stream;
+        kernels::RoiAlignParams params;
+    };
+
+}}} /* namespace cv::dnn::cuda4dnn */
+
+#endif /* OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_ROI_ALIGN_HPP */

modules/dnn/src/cuda4dnn/primitives/scatter_elements.hpp

@@ -0,0 +1,87 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+// Copyright (C) 2026, BigVision LLC, all rights reserved.
+// Third party copyrights are property of their respective owners.
+#ifndef OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_SCATTER_ELEMENTS_HPP
+#define OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_SCATTER_ELEMENTS_HPP
+
+#include "../../op_cuda.hpp"
+
+#include "../csl/stream.hpp"
+
+#include "../kernels/scatter_elements.hpp"
+#include "../kernels/fill_copy.hpp"
+
+#include <opencv2/core.hpp>
+
+#include <cstdint>
+#include <utility>
+#include <vector>
+
+namespace cv { namespace dnn { namespace cuda4dnn {
+
+    template <class T>
+    class ScatterElementsOp final : public CUDABackendNode {
+    public:
+        ScatterElementsOp(csl::Stream stream_, int axis_)
+            : stream(std::move(stream_)), axis(axis_)
+        {
+        }
+
+        void forward(
+            const std::vector<UMat>& inputs,
+            const std::vector<UMat>& outputs,
+            csl::Workspace& workspace) override
+        {
+            CV_UNUSED(workspace);
+            CV_Assert(inputs.size() == 3 && outputs.size() == 1);
+
+            auto data = csl::viewOf<T>(inputs[0]);
+            auto updates = csl::viewOf<T>(inputs[2]);
+            auto output = csl::spanOf<T>(outputs[0]);
+
+            // ScatterElements writes a sparse subset, so the untouched elements come from data
+            if (output.get() != data.get())
+                kernels::copy<T>(stream, output, data);
+
+            const MatShape dataShape = cv::dnn::shape(inputs[0]);
+            const MatShape idxShape = cv::dnn::shape(inputs[1]);
+            const int rank = dataShape.dims;
+            CV_Assert(rank == idxShape.dims && rank >= 1);
+
+            const int axis_ = normalize_axis(axis, rank);
+
+            std::vector<std::int64_t> output_strides(rank);
+            output_strides[rank - 1] = 1;
+            for (int i = rank - 2; i >= 0; --i)
+                output_strides[i] = output_strides[i + 1] * dataShape[i + 1];
+
+            std::vector<std::int64_t> indices_dims(rank);
+            for (int i = 0; i < rank; i++)
+                indices_dims[i] = idxShape[i];
+
+            const std::int64_t axis_dim_size = dataShape[axis_];
+            const int idepth = inputs[1].depth();
+
+            if (idepth == CV_32S) {
+                kernels::scatter_elements<T, std::int32_t>(stream, output,
+                    csl::viewOf<std::int32_t>(inputs[1]), updates,
+                    rank, axis_, indices_dims, output_strides, axis_dim_size);
+            } else if (idepth == CV_64S) {
+                kernels::scatter_elements<T, std::int64_t>(stream, output,
+                    csl::viewOf<std::int64_t>(inputs[1]), updates,
+                    rank, axis_, indices_dims, output_strides, axis_dim_size);
+            } else {
+                CV_Error(Error::StsNotImplemented, "ScatterElements: indices must be CV_32S or CV_64S");
+            }
+        }
+
+    private:
+        csl::Stream stream;
+        int axis;
+    };
+
+}}} /* namespace cv::dnn::cuda4dnn */
+
+#endif /* OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_SCATTER_ELEMENTS_HPP */

modules/dnn/src/cuda4dnn/primitives/scatter_nd.hpp

@@ -0,0 +1,88 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+// Copyright (C) 2026, BigVision LLC, all rights reserved.
+// Third party copyrights are property of their respective owners.
+#ifndef OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_SCATTER_ND_HPP
+#define OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_SCATTER_ND_HPP
+
+#include "../../op_cuda.hpp"
+
+#include "../csl/stream.hpp"
+
+#include "../kernels/scatter_nd.hpp"
+#include "../kernels/fill_copy.hpp"
+
+#include <opencv2/core.hpp>
+
+#include <cstddef>
+#include <cstdint>
+#include <vector>
+#include <utility>
+
+namespace cv { namespace dnn { namespace cuda4dnn {
+
+    template <class T, class TIdx>
+    class ScatterNDOp final : public CUDABackendNode {
+    public:
+        ScatterNDOp(csl::Stream stream_)
+            : stream(std::move(stream_))
+        {
+        }
+
+        void forward(
+            const std::vector<UMat>& inputs,
+            const std::vector<UMat>& outputs,
+            csl::Workspace& workspace) override
+        {
+            CV_UNUSED(workspace);
+            CV_Assert(inputs.size() == 3 && outputs.size() == 1);
+
+            if (inputs[0].total() == 0)
+                return;
+
+            auto data = csl::viewOf<T>(inputs[0]);
+            auto output = csl::spanOf<T>(outputs[0]);
+
+            kernels::copy<T>(stream, output, data);
+
+            if (inputs[1].total() == 0 || inputs[2].total() == 0)
+                return;
+
+            auto indices = csl::viewOf<TIdx>(inputs[1]);
+            auto updates = csl::viewOf<T>(inputs[2]);
+
+            MatShape dataShape = cv::dnn::shape(inputs[0]);
+            MatShape indShape = cv::dnn::shape(inputs[1]);
+            const int r = dataShape.dims;
+            const int q = indShape.dims;
+            const std::int64_t last_index_dimension = indShape[q - 1];
+
+            std::vector<std::int64_t> element_counts(last_index_dimension), input_dims(last_index_dimension);
+            for (std::int64_t j = 0; j < last_index_dimension; j++) {
+                std::int64_t stride = 1;
+                for (int i = static_cast<int>(j) + 1; i < r; i++)
+                    stride *= dataShape[i];
+                element_counts[j] = stride;
+                input_dims[j] = dataShape[static_cast<int>(j)];
+            }
+
+            std::size_t num_indices = 1;
+            for (int i = 0; i < q - 1; i++)
+                num_indices *= static_cast<std::size_t>(indShape[i]);
+
+            std::size_t num_updates_elements = 1;
+            for (int i = static_cast<int>(last_index_dimension); i < r; i++)
+                num_updates_elements *= static_cast<std::size_t>(dataShape[i]);
+
+            kernels::scatter_nd<T, TIdx>(stream, output, indices, updates,
+                                   num_indices, last_index_dimension, element_counts, input_dims, num_updates_elements);
+        }
+
+    private:
+        csl::Stream stream;
+    };
+
+}}} /* namespace cv::dnn::cuda4dnn */
+
+#endif /* OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_SCATTER_ND_HPP */

modules/dnn/src/cuda4dnn/primitives/size.hpp

@@ -0,0 +1,47 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+// Copyright (C) 2026, BigVision LLC, all rights reserved.
+// Third party copyrights are property of their respective owners.
+#ifndef OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_SIZE_HPP
+#define OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_SIZE_HPP
+
+#include "../../op_cuda.hpp"
+
+#include "../csl/stream.hpp"
+
+#include "../kernels/fill_copy.hpp"
+
+#include <opencv2/core.hpp>
+
+#include <cstdint>
+#include <utility>
+
+namespace cv { namespace dnn { namespace cuda4dnn {
+
+    class SizeOp final : public CUDABackendNode {
+    public:
+        SizeOp(csl::Stream stream_) : stream(std::move(stream_)) { }
+
+        void forward(
+            const std::vector<UMat>& inputs,
+            const std::vector<UMat>& outputs,
+            csl::Workspace& workspace) override
+        {
+            CV_UNUSED(workspace);
+            CV_Assert(inputs.size() == 1 && outputs.size() == 1);
+
+            auto output = csl::spanOf<std::int64_t>(outputs[0]);
+            CV_Assert(output.size() == 1);
+
+            const std::int64_t n = static_cast<std::int64_t>(inputs[0].total());
+            kernels::fill<std::int64_t>(stream, output, n);
+        }
+
+    private:
+        csl::Stream stream;
+    };
+
+}}} /* namespace cv::dnn::cuda4dnn */
+
+#endif /* OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_SIZE_HPP */

modules/dnn/src/cuda4dnn/primitives/topk.hpp

@@ -0,0 +1,75 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+// Copyright (C) 2026, BigVision LLC, all rights reserved.
+// Third party copyrights are property of their respective owners.
+#ifndef OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_TOPK_HPP
+#define OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_TOPK_HPP
+
+#include "../../op_cuda.hpp"
+
+#include "../csl/stream.hpp"
+#include "../csl/workspace.hpp"
+
+#include "../kernels/topk.hpp"
+
+#include <opencv2/core.hpp>
+#include <opencv2/dnn/shape_utils.hpp>
+
+#include <cstdint>
+#include <utility>
+#include <vector>
+
+namespace cv { namespace dnn { namespace cuda4dnn {
+
+    template <class T>
+    class TopKOp final : public CUDABackendNode {
+    public:
+        TopKOp(csl::Stream stream_, int axis_, bool largest_, const MatShape& inShape, int k_)
+            : stream(std::move(stream_)), largest(largest_), k(k_)
+        {
+            const int a = normalize_axis(axis_, inShape.dims);
+            CV_Assert(a >= 0 && a < inShape.dims);
+            for (int i = 0; i < a; i++) outer *= inShape[i];
+            for (int i = a + 1; i < inShape.dims; i++) inner *= inShape[i];
+            dim_axis = inShape[a];
+            // the sort path's scratch comes from the shared workspace, sized once per shape
+            if (inner == 1 && static_cast<std::int64_t>(k) * dim_axis > sort_threshold)
+                sort_bytes = kernels::topk_sort_workspace<T>(stream, outer, dim_axis, largest);
+        }
+
+        void forward(
+            const std::vector<UMat>& inputs,
+            const std::vector<UMat>& outputs,
+            csl::Workspace& workspace) override
+        {
+            CV_Assert(inputs.size() >= 1 && outputs.size() == 2);
+
+            auto input = csl::viewOf<T>(inputs[0]);
+            auto values = csl::spanOf<T>(outputs[0]);
+            auto indices = csl::spanOf<std::int64_t>(outputs[1]);
+
+            if (sort_bytes) {
+                auto scratch = csl::WorkspaceAllocator(workspace).get_span<unsigned char>(sort_bytes);
+                kernels::topk_sort<T>(stream, values, indices, input, outer, dim_axis, k, largest,
+                                      scratch.data().get(), sort_bytes);
+                return;
+            }
+
+            kernels::topk<T>(stream, values, indices, input, outer, dim_axis, inner, k, largest);
+        }
+
+        std::size_t get_workspace_memory_in_bytes() const noexcept override { return sort_bytes; }
+
+    private:
+        static constexpr std::int64_t sort_threshold = std::int64_t(1) << 16;
+
+        csl::Stream stream;
+        bool largest;
+        int k, outer = 1, inner = 1, dim_axis = 1;
+        std::size_t sort_bytes = 0;
+    };
+
+}}} /* namespace cv::dnn::cuda4dnn */
+
+#endif /* OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_TOPK_HPP */

modules/dnn/src/cuda4dnn/primitives/trilu.hpp

@@ -0,0 +1,60 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+// Copyright (C) 2026, BigVision LLC, all rights reserved.
+// Third party copyrights are property of their respective owners.
+#ifndef OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_TRILU_HPP
+#define OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_TRILU_HPP
+
+#include "../../op_cuda.hpp"
+
+#include "../csl/stream.hpp"
+
+#include "../kernels/trilu.hpp"
+
+#include <opencv2/core.hpp>
+
+#include <utility>
+#include <vector>
+
+namespace cv { namespace dnn { namespace cuda4dnn {
+
+    template <class T>
+    class TriluOp final : public CUDABackendNode {
+    public:
+        TriluOp(csl::Stream stream_, int k_, bool upper_)
+            : stream(std::move(stream_)), k(k_), upper(upper_) { }
+
+        void forward(
+            const std::vector<UMat>& inputs,
+            const std::vector<UMat>& outputs,
+            csl::Workspace& workspace) override
+        {
+            CV_UNUSED(workspace);
+            CV_Assert(inputs.size() >= 1 && outputs.size() == 1);
+
+            auto input = csl::viewOf<T>(inputs[0]);
+            auto output = csl::spanOf<T>(outputs[0]);
+
+            MatShape inShape = cv::dnn::shape(inputs[0]);
+            const int rank = inShape.dims;
+            CV_Assert(rank >= 1);
+
+            const int width = inShape[rank - 1];
+            const int height = (rank >= 2) ? inShape[rank - 2] : 1;
+
+            int loops = 1;
+            for (int i = 0; i < rank - 2; i++) loops *= inShape[i];
+
+            kernels::trilu<T>(stream, output, input, loops, height, width, k, upper);
+        }
+
+    private:
+        csl::Stream stream;
+        int k;
+        bool upper;
+    };
+
+}}} /* namespace cv::dnn::cuda4dnn */
+
+#endif /* OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_TRILU_HPP */

modules/dnn/src/dnn_common.hpp

@@ -180,6 +180,9 @@ static inline void prindent(std::ostream& strm, int indent)
         strm << ' ';
 }
 
+// Test hook: ops a new-engine net set to DNN_BACKEND_CUDA runs on the CPU, as "name (type)", after the first forward()
+CV_EXPORTS std::vector<std::string> cudaFallbackOps(const Net& net);
+
 CV__DNN_INLINE_NS_END
 
 namespace accessor {

modules/dnn/src/layers/arg_layer.cpp

@@ -4,7 +4,11 @@
 
 #include "../precomp.hpp"
 #include "layers_common.hpp"
-
+#include "../net_impl.hpp"
+#include "../op_cuda.hpp"
+#ifdef HAVE_CUDA
+#include "../cuda4dnn/primitives/arg.hpp"
+#endif
 
 namespace cv { namespace dnn {
 
@@ -43,9 +47,31 @@ public:
 
     virtual bool supportBackend(int backendId) CV_OVERRIDE
     {
+#ifdef HAVE_CUDA
+        if (backendId == DNN_BACKEND_CUDA) {
+            Net::Impl* netimpl_ = getNetImpl(this);
+            if (!netimpl_ || this->inputs.size() != 1)
+                return false;
+            // an intermediate arg has no type until the first forward; refuse only a known unsupported one
+            const int t = netimpl_->argData(this->inputs[0]).type;
+            return t < 0 || isCudaNodeType(t, false);
+        }
+#endif
         return backendId == DNN_BACKEND_OPENCV;
     }
 
+#ifdef HAVE_CUDA
+    Ptr<BackendNode> initCUDA(void* context_,
+                              InputArrayOfArrays inputs_arr,
+                              InputArrayOfArrays) CV_OVERRIDE
+    {
+        auto context = reinterpret_cast<cuda4dnn::csl::CSLContext*>(context_);
+        return make_cuda_node_with_type<cuda4dnn::ArgOp>(preferableTarget, inputs_arr.type(0),
+                                                         std::move(context->stream), axis,
+                                                         op == ArgOp::MAX, select_last_index);
+    }
+#endif
+
     void handleKeepDims(MatShape& shape, const int axis_) const
     {
         if (keepdims)

modules/dnn/src/layers/blank_layer.cpp

@@ -40,6 +40,7 @@
 //
 //M*/
 #include "../precomp.hpp"
+#include "../net_impl.hpp"
 #include "../op_cuda.hpp"
 #include "../op_inf_engine.hpp"
 #include "../ie_ngraph.hpp"
@@ -67,6 +68,17 @@ public:
 #ifdef HAVE_INF_ENGINE
         if (backendId == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH)
             return true;
+#endif
+#ifdef HAVE_CUDA
+        if (backendId == DNN_BACKEND_CUDA) {
+            Net::Impl* netimpl_ = getNetImpl(this);
+            // classic-engine nets carry no graph args to inspect
+            if (!netimpl_ || inputs.empty())
+                return true;
+            // an intermediate arg has no type until the first forward; refuse only a known unsupported one
+            const int t = netimpl_->argData(inputs[0]).type;
+            return t < 0 || isCudaNodeType(t);
+        }
 #endif
         return backendId == DNN_BACKEND_OPENCV ||
                backendId == DNN_BACKEND_CUDA ||
@@ -179,15 +191,16 @@ public:
 #ifdef HAVE_CUDA
     Ptr<BackendNode> initCUDA(
         void *context_,
-        const std::vector<Ptr<BackendWrapper>>& inputs,
-        const std::vector<Ptr<BackendWrapper>>& outputs
+        InputArrayOfArrays inputs_arr,
+        InputArrayOfArrays
     ) override
     {
         auto context = reinterpret_cast<csl::CSLContext*>(context_);
-        if (inputs[0]->getHostMatDepth() == CV_Bool)
+        const int depth = inputs_arr.depth(0);
+        if (depth == CV_Bool)
             return make_cuda_node_bool<cuda4dnn::ReshapeOp>(std::move(context->stream));
         else
-            return make_cuda_node_with_type<cuda4dnn::ReshapeOp>(preferableTarget, inputs[0]->getHostMatDepth(), std::move(context->stream));
+            return make_cuda_node_with_type<cuda4dnn::ReshapeOp>(preferableTarget, depth, std::move(context->stream));
     }
 #endif
 };

modules/dnn/src/layers/cast2_layer.cpp

@@ -115,9 +115,13 @@ public:
             if (inType >= 0 && inType == outType)
                 return isCudaNodeType(inType);
             // an intermediate arg has no type until the first forward; check only the known side(s)
-            auto castable = [](int t) { return t < 0 || t == CV_Bool || t == CV_32S || t == CV_64S || t == CV_32F; };
+            auto castable = [](int t) { return t < 0 || t == CV_Bool || t == CV_32S || t == CV_64S || t == CV_32F ||
+                                               t == CV_8U || t == CV_8S; };
             if (!castable(inType) || !castable(outType))
                 return false;
+            // an unknown output type still has to match the static 'to' attribute
+            if (outType < 0 && toCvDepth_ >= 0 && !castable(resolveStorageDepth(toCvDepth_, false)))
+                return false;
             // the FP16 target stores float as half, which has no CastOp instantiation; an unknown type may be float
             return netimpl_->preferableTarget != DNN_TARGET_CUDA_FP16 ||
                    (inType >= 0 && outType >= 0 && inType != CV_32F && outType != CV_32F);
@@ -147,6 +151,8 @@ public:
         case CV_32S: return makeCastFrom<int32_t>(outDepth, std::move(context->stream));
         case CV_64S: return makeCastFrom<int64_t>(outDepth, std::move(context->stream));
         case CV_32F: return makeCastFrom<float>(outDepth, std::move(context->stream));
+        case CV_8U: return makeCastFrom<uint8_t>(outDepth, std::move(context->stream));
+        case CV_8S: return makeCastFrom<int8_t>(outDepth, std::move(context->stream));
         }
         CV_Error(Error::BadDepth, "Cast: unsupported input type on CUDA");
         return Ptr<BackendNode>();
@@ -161,6 +167,8 @@ public:
         case CV_32S: return Ptr<BackendNode>(new cuda4dnn::CastOp<int32_t, TIn>(std::move(stream)));
         case CV_64S: return Ptr<BackendNode>(new cuda4dnn::CastOp<int64_t, TIn>(std::move(stream)));
         case CV_32F: return Ptr<BackendNode>(new cuda4dnn::CastOp<float, TIn>(std::move(stream)));
+        case CV_8U: return Ptr<BackendNode>(new cuda4dnn::CastOp<uint8_t, TIn>(std::move(stream)));
+        case CV_8S: return Ptr<BackendNode>(new cuda4dnn::CastOp<int8_t, TIn>(std::move(stream)));
         }
         CV_Error(Error::BadDepth, "Cast: unsupported output type on CUDA");
         return Ptr<BackendNode>();

modules/dnn/src/layers/clip_layer.cpp

@@ -24,8 +24,14 @@ static inline void clampFloatChunkDispatch(const float* src, float* dst,
 #undef CV_CPU_DISPATCH_MODES_ALL
 
 #include "layers_common.hpp"
+#include "../net_impl.hpp"
 #include <opencv2/dnn/shape_utils.hpp>
 #include <opencv2/core/hal/interface.h>
+
+#ifdef HAVE_CUDA
+#include "../op_cuda.hpp"
+#include "../cuda4dnn/primitives/activation.hpp"
+#endif
 #include <limits>
 #include <cfloat>
 #include <cmath>
@@ -113,9 +119,51 @@ public:
 
     virtual bool supportBackend(int backendId) CV_OVERRIDE
     {
+#ifdef HAVE_CUDA
+        if (backendId == DNN_BACKEND_CUDA) {
+            Net::Impl* netimpl_ = getNetImpl(this);
+            if (!netimpl_ || inputs.empty() || inputs.size() > 3)
+                return false;
+            // initCUDA reads min/max on the host
+            for (size_t i = 1; i < inputs.size(); i++)
+                if (!inputs[i].empty() && !netimpl_->isConstArg(inputs[i]))
+                    return false;
+            // an intermediate arg has no type until the first forward; refuse only a known unsupported one
+            const int t = netimpl_->argData(inputs[0]).type;
+            return t < 0 || isCudaNodeType(t, false);
+        }
+#endif
         return backendId == DNN_BACKEND_OPENCV;
     }
 
+#ifdef HAVE_CUDA
+    Ptr<BackendNode> initCUDA(void* context_, InputArrayOfArrays inputs_arr, InputArrayOfArrays) CV_OVERRIDE
+    {
+        auto context = reinterpret_cast<cuda4dnn::csl::CSLContext*>(context_);
+        Net::Impl* netimpl_ = getNetImpl(this);
+        CV_Assert(netimpl_);
+        const int type = inputs_arr.type(0);
+        // a half device tensor holds FP32 data, so it takes the float range
+        const int depth = CV_MAT_DEPTH(type) == CV_16F ? CV_32F : CV_MAT_DEPTH(type);
+
+        // same bounds as forward(): const inputs win over attributes, then the type's range
+        auto constScalar = [&](size_t i, double& v) {
+            if (inputs.size() <= i || inputs[i].empty())
+                return false;
+            Mat tmp;
+            netimpl_->argTensor(inputs[i]).convertTo(tmp, CV_64F);
+            CV_Assert(tmp.total() == 1);
+            v = tmp.at<double>(0);
+            return true;
+        };
+        double lo = hasMin ? minValue : typeMin(depth), hi = hasMax ? maxValue : typeMax(depth);
+        constScalar(1, lo);
+        constScalar(2, hi);
+        CV_Assert(lo <= hi);
+        return make_cuda_node_with_type<cuda4dnn::ClipOp>(preferableTarget, type, std::move(context->stream), lo, hi);
+    }
+#endif
+
     // Clip rewrites values in place of the input layout.
     virtual bool alwaysSupportInplace() const CV_OVERRIDE { return true; }
 

modules/dnn/src/layers/conv2_layer.cpp

@@ -16,6 +16,7 @@
 #include <opencv2/dnn/layer.details.hpp>  // CV_DNN_REGISTER_EXEC_CLASS
 #ifdef HAVE_CUDA
 #include "../cuda4dnn/primitives/convolution.hpp"
+#include "fusion_cuda.hpp"  // FusedPostOpNode and the activation ops
 using namespace cv::dnn::cuda4dnn;
 #endif
 
@@ -547,6 +548,20 @@ public:
         return absorbAffine(expr) || absorbKernel(expr);
     }
 
+    //! The kernel activations ConvolutionOp applies itself.
+    static bool cudaConvKernel(ActivationFunc fn)
+    {
+        return fn == getActivationFunc(ACTIV_SIGMOID) || fn == getActivationFunc(ACTIV_SWISH) ||
+               fn == getActivationFunc(ACTIV_MISH) || fn == getActivationFunc(ACTIV_TANH);
+    }
+
+    //! Kernel activations the CUDA path runs as a separate op after the convolution.
+    static bool cudaPostKernel(ActivationFunc fn)
+    {
+        return fn == getActivationFunc(ACTIV_HARDSWISH) || fn == getActivationFunc(ACTIV_GELU) ||
+               fn == getActivationFunc(ACTIV_GELU_APPROX) || fn == getActivationFunc(ACTIV_EXP);
+    }
+
     //! Takes an expression the layer has a kernel for. Only a one-layer chain carries
     //! one, so this cannot swallow math absorbAffine() just turned down.
     bool absorbKernel(const Ptr<AdjacencyGraph>& expr)
@@ -1199,13 +1214,12 @@ public:
             return false;
         if (!activ.empty())
             return false;
-        // a fused kernel activation runs on CUDA only if ConvolutionOp has the same one
-        if (activationFunc != nullptr &&
-            activationFunc != getActivationFunc(ACTIV_SIGMOID) && activationFunc != getActivationFunc(ACTIV_SWISH) &&
-            activationFunc != getActivationFunc(ACTIV_MISH) && activationFunc != getActivationFunc(ACTIV_TANH))
+        // a fused kernel activation runs on CUDA inside ConvolutionOp or as a post-op after it
+        if (activationFunc != nullptr && !cudaConvKernel(activationFunc) && !cudaPostKernel(activationFunc))
             return false;
         if (fastActivation != FAST_ACTIV_NONE && fastActivation != FAST_ACTIV_RELU &&
-            fastActivation != FAST_ACTIV_LEAKY_RELU && fastActivation != FAST_ACTIV_CLIP)
+            fastActivation != FAST_ACTIV_LEAKY_RELU && fastActivation != FAST_ACTIV_CLIP &&
+            fastActivation != FAST_ACTIV_PRELU)
             return false;
         return true;
     }
@@ -1256,11 +1270,28 @@ public:
             biasMat = fusedBias;  // already b*scale + bn_bias
         }
 
+        // activations ConvolutionOp lacks run as their own op on the convolution's output
+        Ptr<BackendNode> post;
+        if (fastActivation == FAST_ACTIV_PRELU) {
+            Mat slope(1, (int)activParams.size(), CV_32F, activParams.data());
+            post = make_cuda_node<cuda4dnn::ChannelwiseReLUOp>(targetId, context.stream, slope.clone());
+        } else if (activationFunc == getActivationFunc(ACTIV_HARDSWISH)) {
+            post = make_cuda_node<cuda4dnn::HardSwishOp>(targetId, context.stream);
+        } else if (activationFunc == getActivationFunc(ACTIV_GELU)) {
+            post = make_cuda_node<cuda4dnn::GeluOp>(targetId, context.stream);
+        } else if (activationFunc == getActivationFunc(ACTIV_GELU_APPROX)) {
+            post = make_cuda_node<cuda4dnn::GeluApproximationOp>(targetId, context.stream);
+        } else if (activationFunc == getActivationFunc(ACTIV_EXP)) {
+            const float scale = activParams.size() > 0 ? activParams[0] : 1.f;
+            const float shift = activParams.size() > 1 ? activParams[1] : 0.f;
+            post = make_cuda_node<cuda4dnn::ExpOp>(targetId, context.stream, scale, shift);
+        }
+
         config.activation_type = ConvolutionConfiguration::ActivationType::IDENTITY;
         config.relu_negative_slope = 0.f;
         config.crelu_floor = 0.f; config.crelu_ceil = 0.f;
         config.power_exp = 1.f; config.power_scale = 1.f; config.power_shift = 0.f;
-        bool hasAct = fastActivation != FAST_ACTIV_NONE || activationFunc != nullptr;
+        bool hasAct = (fastActivation != FAST_ACTIV_NONE || activationFunc != nullptr) && post.empty();
         if (activationFunc == getActivationFunc(ACTIV_SIGMOID)) {
             config.activation_type = ConvolutionConfiguration::ActivationType::SIGMOID;
         } else if (activationFunc == getActivationFunc(ACTIV_SWISH)) {
@@ -1289,9 +1320,12 @@ public:
         else
             config.fusion_mode = ConvolutionConfiguration::FusionMode::NONE;
 
-        return make_cuda_node<cuda4dnn::ConvolutionOp>(
+        Ptr<BackendNode> convNode = make_cuda_node<cuda4dnn::ConvolutionOp>(
             targetId, std::move(context.stream), std::move(context.cudnn_handle),
             config, filters, biasMat);
+        if (post.empty())
+            return convNode;
+        return makePtr<FusedPostOpNode>(convNode.dynamicCast<CUDABackendNode>(), post.dynamicCast<CUDABackendNode>());
     }
 #endif
 

modules/dnn/src/layers/convtranspose_layer.cpp

@@ -12,6 +12,11 @@
 #include "opencv2/core/hal/intrin.hpp"
 #include <opencv2/core/utils/configuration.private.hpp>
 
+#ifdef HAVE_CUDA
+#include "../op_cuda.hpp"
+#include "../cuda4dnn/primitives/conv_transpose2.hpp"
+#endif
+
 namespace cv
 {
 namespace dnn
@@ -96,6 +101,86 @@ public:
         return inptype0;
     }
 
+    // compute actual pads for SAME/VALID auto-padding, or from an explicit output_shape
+    std::vector<int> resolvePads(const MatShape& inpshape, const MatShape& outshape,
+                                 const MatShape& wshape, int nsd) const
+    {
+        std::vector<int> pads_resolved = pads;
+        if (auto_pad != AUTO_PAD_NONE || !explicit_out_shape.empty()) {
+            pads_resolved.resize(nsd * 2, 0);
+            for (int i = 0; i < nsd; i++) {
+                int inpsz   = inpshape[2 + i];
+                int outsz   = outshape[2 + i];
+                int adj_i   = adjust_pads.empty() ? 0 : adjust_pads[i];
+                int stride  = strides.empty() ? 1 : strides[i];
+                int dil     = dilations.empty() ? 1 : dilations[i];
+                int ki      = wshape[2 + i];
+                int total   = (inpsz - 1) * stride + dil * (ki - 1) + 1 + adj_i - outsz;
+                total = std::max(total, 0);
+                // ONNX splits total_padding evenly; odd pixel goes last for SAME_UPPER.
+                int pb = (auto_pad == AUTO_PAD_SAME_UPPER) ? total / 2 : total - total / 2;
+                pads_resolved[i]       = pb;
+                pads_resolved[nsd + i] = total - pb;
+            }
+        }
+        return pads_resolved;
+    }
+
+    virtual bool supportBackend(int backendId) CV_OVERRIDE
+    {
+#ifdef HAVE_CUDA
+        if (backendId == DNN_BACKEND_CUDA) {
+            Net::Impl* netimpl_ = getNetImpl(this);
+            // weights and bias come as device inputs, or packed into the layer (constant ones)
+            const bool packed = inputs.size() == 1 && !origWeights.empty();
+            if (!netimpl_ || (!packed && (inputs.size() < 2 || inputs.size() > 3)))
+                return false;
+            // 2-D and FP32 only; the FP16 target keeps the CPU path
+            if (netimpl_->preferableTarget == DNN_TARGET_CUDA_FP16)
+                return false;
+            const int xdims = netimpl_->argData(inputs[0]).shape.dims;
+            const int wdims = packed ? origWeights.dims : netimpl_->argData(inputs[1]).shape.dims;
+            if ((xdims >= 0 && xdims != 4) || (wdims >= 0 && wdims != 4))
+                return false;
+            // an intermediate arg has no type until the first forward; refuse only a known unsupported one
+            for (const Arg& a : inputs) {
+                const int t = a.empty() ? -1 : netimpl_->argData(a).type;
+                if (t >= 0 && t != CV_32F)
+                    return false;
+            }
+            return true;
+        }
+#endif
+        return backendId == DNN_BACKEND_OPENCV;
+    }
+
+#ifdef HAVE_CUDA
+    Ptr<BackendNode> initCUDA(void* context_, InputArrayOfArrays inputs_arr, InputArrayOfArrays outputs_arr) CV_OVERRIDE
+    {
+        auto context = reinterpret_cast<cuda4dnn::csl::CSLContext*>(context_);
+        const int nsd = 2;
+
+        const bool packed = inputs_arr.total() == 1;
+        cuda4dnn::ConvTranspose2Configuration config;
+        config.input_shape = inputs_arr.shape(0);
+        config.weight_shape = packed ? origWeights.shape() : inputs_arr.shape(1);
+        config.output_shape = outputs_arr.shape(0);
+        config.groups = ngroups;
+        // same pads as the CPU path; output_padding only grows the end, which the output shape carries
+        std::vector<int> pads_resolved = resolvePads(config.input_shape, config.output_shape, config.weight_shape, nsd);
+        for (int i = 0; i < nsd; i++) {
+            config.strides.push_back(strides.empty() ? 1 : strides[i]);
+            config.dilations.push_back(dilations.empty() ? 1 : dilations[i]);
+            config.pads_begin.push_back(pads_resolved.empty() ? 0 : pads_resolved[i]);
+        }
+        if (packed)
+            return make_cuda_node<cuda4dnn::ConvTranspose2Op>(preferableTarget, std::move(context->stream),
+                                                              std::move(context->cublas_handle), config, origWeights, bias);
+        return make_cuda_node<cuda4dnn::ConvTranspose2Op>(preferableTarget, std::move(context->stream),
+                                                          std::move(context->cublas_handle), config);
+    }
+#endif
+
     virtual void setWeights(InputArray weights_arr, InputArray bias_arr,
                             int C0, int /*accuracy*/) CV_OVERRIDE
     {
@@ -110,6 +195,7 @@ public:
             rawWeights.convertTo(wfloat, CV_32F);
         else
             wfloat = rawWeights;
+        origWeights = wfloat.clone();  // the unpacked [Cin, Cout/g, kh, kw] filter the CUDA path reads
         packGemmWeights(wfloat, C0);
         weights.release();
         if (gemmWeights.empty())
@@ -518,26 +604,8 @@ public:
                                              pads, adjust_pads, auto_pad);
         applyExplicitOutShape(outshape);
 
-        // compute actual pads for SAME/VALID auto-padding, or from an explicit output_shape
-        int nsd = inpshape.dims - 3;
-        std::vector<int> pads_resolved = pads;
-        if (auto_pad != AUTO_PAD_NONE || !explicit_out_shape.empty()) {
-            pads_resolved.resize(nsd * 2, 0);
-            for (int i = 0; i < nsd; i++) {
-                int inpsz   = inpshape[2 + i];
-                int outsz   = outshape[2 + i];
-                int adj_i   = adjust_pads.empty() ? 0 : adjust_pads[i];
-                int stride  = strides.empty() ? 1 : strides[i];
-                int dil     = dilations.empty() ? 1 : dilations[i];
-                int ki      = wshape0[2 + i];
-                int total   = (inpsz - 1) * stride + dil * (ki - 1) + 1 + adj_i - outsz;
-                total = std::max(total, 0);
-                // ONNX splits total_padding evenly; odd pixel goes last for SAME_UPPER.
-                int pb = (auto_pad == AUTO_PAD_SAME_UPPER) ? total / 2 : total - total / 2;
-                pads_resolved[i]       = pb;
-                pads_resolved[nsd + i] = total - pb;
-            }
-        }
+        // block layout: N, C/C0, spatial..., C0
+        std::vector<int> pads_resolved = resolvePads(inpshape, outshape, wshape0, inpshape.dims - 3);
 
         int outtype = inferType(inptype);
 
@@ -594,6 +662,7 @@ public:
     std::vector<int> emptyKernelShape;
     std::vector<int> explicit_out_shape;
     Mat weights, bias;
+    Mat origWeights;  // unpacked FP32 filter from setWeights(), kept for the CUDA path
     Mat gemmWeights;  // MLAS-packed B per (group, channel chunk), empty without MLAS
     size_t gemmChunkStep = 0;
     int gemmCg = 0, gemmKg = 0, gemmKgp = 0, gemmKC = 0, gemmChunks = 0, gemmKsize = 0, gemmC0 = 0;

modules/dnn/src/layers/cumprod_layer.cpp

@@ -6,6 +6,11 @@
 
 #include "../precomp.hpp"
 #include "layers_common.hpp"
+#include "../net_impl.hpp"
+#include "../op_cuda.hpp"
+#ifdef HAVE_CUDA
+#include "../cuda4dnn/primitives/cum_scan.hpp"
+#endif
 
 #include <opencv2/dnn/shape_utils.hpp>
 
@@ -29,7 +34,44 @@ public:
         setParamsFrom(params);
     }
 
-    bool supportBackend(int backendId) CV_OVERRIDE { return backendId == DNN_BACKEND_OPENCV; }
+    bool supportBackend(int backendId) CV_OVERRIDE
+    {
+#ifdef HAVE_CUDA
+        if (backendId == DNN_BACKEND_CUDA) {
+            Net::Impl* netimpl_ = getNetImpl(this);
+            if (!netimpl_ || this->inputs.empty())
+                return false;
+            // the axis is baked into the node, so it must be a constant
+            if (this->inputs.size() > 1 && !netimpl_->isConstArg(this->inputs[1]))
+                return false;
+            // an intermediate arg has no type until the first forward; refuse only a known unsupported one
+            const int t = netimpl_->argData(this->inputs[0]).type;
+            return t < 0 || t == CV_32F || t == CV_16F || t == CV_32S || t == CV_64S;
+        }
+#endif
+        return backendId == DNN_BACKEND_OPENCV;
+    }
+
+#ifdef HAVE_CUDA
+    Ptr<BackendNode> initCUDA(void* context_,
+                              InputArrayOfArrays inputs_arr,
+                              InputArrayOfArrays) CV_OVERRIDE
+    {
+        auto context = reinterpret_cast<cuda4dnn::csl::CSLContext*>(context_);
+        int axis = axis_raw;
+        if (this->inputs.size() > 1) {
+            Net::Impl* netimpl_ = getNetImpl(this);
+            CV_Assert(netimpl_ && netimpl_->isConstArg(this->inputs[1]));
+            std::vector<int> av;
+            tensorToIntVec(netimpl_->argTensor(this->inputs[1]), av);
+            CV_Assert(av.size() == 1);
+            axis = av[0];
+        }
+        return make_cuda_node_with_type<cuda4dnn::CumScanOp>(preferableTarget, inputs_arr.type(0),
+                                                             std::move(context->stream), axis, true,
+                                                             exclusive_raw == 1, reverse_raw == 1);
+    }
+#endif
 
     bool getMemoryShapes(const std::vector<MatShape>& inputs, const int,
                          std::vector<MatShape>& outputs, std::vector<MatShape>&) const CV_OVERRIDE

modules/dnn/src/layers/cumsum_layer.cpp

@@ -6,6 +6,11 @@
 #include "../op_inf_engine.hpp"
 #include "../ie_ngraph.hpp"
 #include "layers_common.hpp"
+#include "../net_impl.hpp"
+#include "../op_cuda.hpp"
+#ifdef HAVE_CUDA
+#include "../cuda4dnn/primitives/cum_scan.hpp"
+#endif
 
 #include <opencv2/dnn/shape_utils.hpp>
 
@@ -27,10 +32,44 @@ public:
 
     virtual bool supportBackend(int backendId) CV_OVERRIDE
     {
+#ifdef HAVE_CUDA
+        if (backendId == DNN_BACKEND_CUDA) {
+            Net::Impl* netimpl_ = getNetImpl(this);
+            if (!netimpl_ || this->inputs.empty())
+                return false;
+            // the axis is baked into the node, so it must be a constant
+            if (this->inputs.size() > 1 && !netimpl_->isConstArg(this->inputs[1]))
+                return false;
+            // an intermediate arg has no type until the first forward; refuse only a known unsupported one
+            const int t = netimpl_->argData(this->inputs[0]).type;
+            return t < 0 || t == CV_32F || t == CV_16F || t == CV_32S || t == CV_64S;
+        }
+#endif
         return backendId == DNN_BACKEND_OPENCV ||
                backendId == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH;
     }
 
+#ifdef HAVE_CUDA
+    Ptr<BackendNode> initCUDA(void* context_,
+                              InputArrayOfArrays inputs_arr,
+                              InputArrayOfArrays) CV_OVERRIDE
+    {
+        auto context = reinterpret_cast<cuda4dnn::csl::CSLContext*>(context_);
+        int axis = axis_raw;
+        if (this->inputs.size() > 1) {
+            Net::Impl* netimpl_ = getNetImpl(this);
+            CV_Assert(netimpl_ && netimpl_->isConstArg(this->inputs[1]));
+            std::vector<int> av;
+            tensorToIntVec(netimpl_->argTensor(this->inputs[1]), av);
+            CV_Assert(av.size() == 1);
+            axis = av[0];
+        }
+        return make_cuda_node_with_type<cuda4dnn::CumScanOp>(preferableTarget, inputs_arr.type(0),
+                                                             std::move(context->stream), axis, false,
+                                                             exclusive_raw == 1, reverse_raw == 1);
+    }
+#endif
+
     bool getMemoryShapes(const std::vector<MatShape> &inputs,
                          const int requiredOutputs,
                          std::vector<MatShape> &outputs,