feat(dnn,src,imperative): add groupnorm op

GitOrigin-RevId: de3c3d10e5

ci/compatibility/fbs/V2-backup/opr_param_defs.fbs

@@ -1869,6 +1869,13 @@ table LayerNorm {
     normalized_size:ulong = 1;
 }
 
+table GroupNorm {
+    affine:bool = true;
+    eps:float = 1e-5;
+    group:uint = 1;
+    format:ConvolutionFormat = NCHW;
+}
+
 table Dropout {
     drop_prob:float = 0;
     seed:ulong = 0;

ci/compatibility/fbs/V2-backup/schema_v2.fbs

@@ -140,6 +140,7 @@ union OperatorParam {
     param.LSTM = 89,
     param.Softmax = 90,
     param.Diag = 91,
+    param.GroupNorm = 92,
 }
 
 table Operator {

dnn/include/megdnn/oprs/nn.h

@@ -2430,6 +2430,76 @@ protected:
             const TensorLayout& dhx, const TensorLayout& dcx, const TensorLayout& dw,
             size_t workspace_in_bytes);
 };
+
+class GroupNormBase : public OperatorBase {
+    DEF_OPR_IMPL_CTOR(GroupNormBase, OperatorBase);
+    DEF_OPR_PARAM(GroupNorm);
+
+protected:
+    void deduce_layout_fwd(
+            const TensorLayout& data, const TensorLayout& weight,
+            const TensorLayout& bias, TensorLayout& dst, TensorLayout& mean,
+            TensorLayout& rstd);
+    void check_layout_fwd(
+            const TensorLayout& data, const TensorLayout& weight,
+            const TensorLayout& bias, const TensorLayout& dst, const TensorLayout& mean,
+            const TensorLayout& rstd);
+};
+
+class GroupNormForward : public GroupNormBase {
+    DEF_OPR_IMPL(GroupNormForward, GroupNormBase, 3, 3);
+
+public:
+    virtual void exec(
+            _megdnn_tensor_in data, _megdnn_tensor_in weight, _megdnn_tensor_in bias,
+            _megdnn_tensor_out dst, _megdnn_tensor_out mean, _megdnn_tensor_out rstd,
+            _megdnn_workspace workspace) = 0;
+    MGE_WIN_DECLSPEC_FUC void deduce_layout(
+            const TensorLayout& data, const TensorLayout& weight,
+            const TensorLayout& bias, TensorLayout& dst, TensorLayout& mean,
+            TensorLayout& rstd);
+    virtual size_t get_workspace_in_bytes(
+            const TensorLayout& data, const TensorLayout& weight,
+            const TensorLayout& bias, const TensorLayout& dst, const TensorLayout& mean,
+            const TensorLayout& rstd) = 0;
+
+protected:
+    void check_exec(
+            const TensorLayout& data, const TensorLayout& weight,
+            const TensorLayout& bias, const TensorLayout& dst, const TensorLayout& mean,
+            const TensorLayout& rstd, size_t workspace_in_bytes);
+};
+using GroupNorm = GroupNormForward;
+
+class GroupNormBackward : public GroupNormBase {
+    DEF_OPR_IMPL(GroupNormBackward, GroupNormBase, 5, 3);
+
+public:
+    virtual void exec(
+            _megdnn_tensor_in diff, _megdnn_tensor_in data, _megdnn_tensor_in weight,
+            _megdnn_tensor_in mean, _megdnn_tensor_in rstd, _megdnn_tensor_out ddata,
+            _megdnn_tensor_out dweight, _megdnn_tensor_out dbias,
+            _megdnn_workspace workspace) = 0;
+    void deduce_layout(
+            const TensorLayout& diff, const TensorLayout& data,
+            const TensorLayout& weight, const TensorLayout& mean,
+            const TensorLayout& rstd, TensorLayout& ddata, TensorLayout& dweight,
+            TensorLayout& dbias);
+    virtual size_t get_workspace_in_bytes(
+            const TensorLayout& diff, const TensorLayout& data,
+            const TensorLayout& weight, const TensorLayout& mean,
+            const TensorLayout& rstd, const TensorLayout& ddata,
+            const TensorLayout& dweight, const TensorLayout& dbias) = 0;
+
+protected:
+    void check_exec(
+            const TensorLayout& diff, const TensorLayout& data,
+            const TensorLayout& weight, const TensorLayout& mean,
+            const TensorLayout& rstd, const TensorLayout& ddata,
+            const TensorLayout& dweight, const TensorLayout& dbias,
+            size_t workspace_in_bytes);
+};
+
 }  // namespace megdnn
 #include "megdnn/internal/opr_header_epilogue.h"
 

dnn/scripts/opr_param_defs.py

@@ -1247,6 +1247,13 @@ PADDING_MODES = [Doc('REPLICATE = 0', 'aaaaaa|abcdefgh|hhhhhhh'),
  .add_fields('uint64', 'normalized_size', '1')
 )
 
+(pdef('GroupNorm')
+ .add_fields('bool', 'affine', 'true')
+ .add_fields('float32', 'eps', '1e-5f')
+ .add_fields('uint32', 'group', '1')
+ .add_enum_alias('Format', 'Convolution')
+)
+
 (pdef('Dropout')
  .add_fields('float32', 'drop_prob', '0')
  .add_fields('uint64', 'seed', '0')

dnn/src/common/group_norm.cpp

@@ -0,0 +1,121 @@
+#include "megdnn/oprs.h"
+
+#include "src/common/utils.h"
+
+namespace megdnn {
+
+using Param = GroupNormBase::Param;
+
+void GroupNormBase::deduce_layout_fwd(
+        const TensorLayout& data, const TensorLayout& weight, const TensorLayout& bias,
+        TensorLayout& dst, TensorLayout& mean, TensorLayout& rstd) {
+    MEGDNN_MARK_USED_VAR(weight);
+    MEGDNN_MARK_USED_VAR(bias);
+    size_t N = data.shape[0];
+    size_t group = param().group;
+    TensorLayout unnormalized_layout({N, group}, dtype::Float32());
+    dst = data;
+    mean = unnormalized_layout;
+    rstd = unnormalized_layout;
+}
+
+void GroupNormBase::check_layout_fwd(
+        const TensorLayout& data, const TensorLayout& weight, const TensorLayout& bias,
+        const TensorLayout& dst, const TensorLayout& mean, const TensorLayout& rstd) {
+    megdnn_assert_contiguous(data);
+    megdnn_assert_contiguous(weight);
+    megdnn_assert_contiguous(bias);
+    megdnn_assert_contiguous(dst);
+    megdnn_assert_contiguous(mean);
+    megdnn_assert_contiguous(rstd);
+    auto errmsg = [&]() {
+        return megdnn_layout_msg(data) + ", " + megdnn_layout_msg(weight) + ", " +
+               megdnn_layout_msg(bias) + ", " + megdnn_layout_msg(dst) + ", " +
+               megdnn_layout_msg(mean) + ", " + megdnn_layout_msg(rstd);
+    };
+    MEGDNN_MARK_USED_VAR(errmsg);
+
+    megdnn_assert(data.eq_layout(dst), "%s", errmsg().c_str());
+    megdnn_assert(weight.eq_layout(bias), "%s", errmsg().c_str());
+    megdnn_assert(mean.eq_layout(rstd), "%s", errmsg().c_str());
+
+    auto p = param();
+    size_t C = data.shape[1];
+    size_t group = p.group;
+    megdnn_assert(
+            group > 0, "Expected num groups to be greater than 0, got %zu", group);
+    megdnn_assert(
+            C % group == 0,
+            "Expected number of channels in input to be divisible by num_groups, but "
+            "got Channel of shape %zu and num_groups= %zu",
+            C, group);
+}
+
+void GroupNormForward::deduce_layout(
+        const TensorLayout& data, const TensorLayout& weight, const TensorLayout& bias,
+        TensorLayout& dst, TensorLayout& mean, TensorLayout& rstd) {
+    deduce_layout_fwd(data, weight, bias, dst, mean, rstd);
+}
+
+void GroupNormForward::check_exec(
+        const TensorLayout& data, const TensorLayout& weight, const TensorLayout& bias,
+        const TensorLayout& dst, const TensorLayout& mean, const TensorLayout& rstd,
+        size_t workspace_in_bytes) {
+    check_layout_fwd(data, weight, bias, dst, mean, rstd);
+    auto required_workspace_in_bytes =
+            get_workspace_in_bytes(data, weight, bias, dst, mean, rstd);
+    megdnn_assert(workspace_in_bytes >= required_workspace_in_bytes);
+}
+
+void GroupNormBackward::deduce_layout(
+        const TensorLayout& diff, const TensorLayout& data, const TensorLayout& weight,
+        const TensorLayout& mean, const TensorLayout& rstd, TensorLayout& ddata,
+        TensorLayout& dweight, TensorLayout& dbias) {
+    MEGDNN_MARK_USED_VAR(diff);
+    MEGDNN_MARK_USED_VAR(mean);
+    MEGDNN_MARK_USED_VAR(rstd);
+    ddata = data;
+    dweight = weight;
+    dbias = weight;
+}
+
+void GroupNormBackward::check_exec(
+        const TensorLayout& diff, const TensorLayout& data, const TensorLayout& weight,
+        const TensorLayout& mean, const TensorLayout& rstd, const TensorLayout& ddata,
+        const TensorLayout& dweight, const TensorLayout& dbias,
+        size_t workspace_in_bytes) {
+    auto p = param();
+    auto required_workspace_in_bytes = get_workspace_in_bytes(
+            diff, data, weight, mean, rstd, ddata, dweight, dbias);
+    megdnn_assert(workspace_in_bytes >= required_workspace_in_bytes);
+
+    megdnn_assert_contiguous(diff);
+    megdnn_assert_contiguous(data);
+    megdnn_assert_contiguous(mean);
+    megdnn_assert_contiguous(rstd);
+    megdnn_assert_contiguous(ddata);
+    if (p.affine) {
+        megdnn_assert_contiguous(weight);
+        megdnn_assert_contiguous(dweight);
+        megdnn_assert_contiguous(dbias);
+    }
+
+    auto errmsg = [&]() {
+        return megdnn_layout_msg(diff) + ", " + megdnn_layout_msg(data) + ", " +
+               megdnn_layout_msg(weight) + ", " + megdnn_layout_msg(mean) + ", " +
+               megdnn_layout_msg(rstd) + ", " + megdnn_layout_msg(ddata) + ", " +
+               megdnn_layout_msg(dweight) + ", " + megdnn_layout_msg(dbias);
+    };
+    MEGDNN_MARK_USED_VAR(errmsg);
+
+    megdnn_assert(data.eq_layout(ddata), "%s", errmsg().c_str());
+    megdnn_assert(mean.eq_layout(rstd), "%s", errmsg().c_str());
+    if (p.affine) {
+        megdnn_assert(weight.eq_layout(dweight), "%s", errmsg().c_str());
+        megdnn_assert(weight.eq_layout(dbias), "%s", errmsg().c_str());
+    }
+}
+
+}  // namespace megdnn
+
+// vim: syntax=cpp.doxygen

dnn/src/common/handle_impl.h

@@ -216,7 +216,9 @@ private:
     cb(NormForward) \
     cb(RegionRestrictedConvolutionForward) \
     cb(RegionRestrictedConvolutionBackwardData) \
-    cb(RegionRestrictedConvolutionBackwardFilter)
+    cb(RegionRestrictedConvolutionBackwardFilter) \
+    cb(GroupNormForward) \
+    cb(GroupNormBackward)
 // clang-format on
 
 /*!

dnn/src/common/opr_trait.h

@@ -142,6 +142,8 @@ DEF(SoftmaxBackward, 3, true, false);
 DEF(RegionRestrictedConvolutionForward, 5, true, true);
 DEF(RegionRestrictedConvolutionBackwardData, 5, true, false);
 DEF(RegionRestrictedConvolutionBackwardFilter, 5, true, false);
+DEF(GroupNormForward, 6, true, true);
+DEF(GroupNormBackward, 8, true, true);
 }  // namespace megdnn
 
 // vim: syntax=cpp.doxygen

dnn/src/cuda/group_norm/group_norm_cuda.cu

@@ -0,0 +1,529 @@
+#include <stdio.h>
+#include <thrust/pair.h>
+#include <thrust/tuple.h>
+#include <cfloat>
+#include "megdnn/arch.h"
+#include "megdnn/basic_types.h"
+#include "megdnn/dtype.h"
+#include "src/cuda/cuda_shfl_compat.cuh"
+#include "src/cuda/group_norm/group_norm_cuda.cuh"
+#include "src/cuda/utils.cuh"
+
+namespace megdnn {
+namespace cuda {
+namespace group_norm {
+
+// warp size may be used as array length, or used in host function,
+// so we define WARP_SIZE rather than using warpSize
+#define WARP_SIZE 32
+
+template <size_t kStart, size_t kEnd, bool kStop>
+struct Compare {
+    template <typename T>
+    __host__ __device__ inline static bool Run(const T* d1, const T* d2) {
+        return d1[kStart] == d2[kStart] &&
+               Compare<kStart + 1, kEnd, kStart + 1 == kEnd>::Run(d1, d2);
+    }
+};
+
+template <size_t kStart, size_t kEnd>
+struct Compare<kStart, kEnd, true> {
+    template <typename T>
+    __host__ __device__ inline constexpr static bool Run(const T* d1, const T* d2) {
+        return true;
+    }
+};
+
+template <size_t N>
+using UnrollCompare = Compare<0, N, N == 0>;
+
+template <typename T, size_t kStart, size_t kEnd, bool kStop>
+struct UnrollVarArgsAssignImpl {
+    template <typename... Args>
+    __host__ __device__ inline static void Run(T* d, T val, Args... args) {
+        static_assert(sizeof...(args) + 1 == kEnd - kStart, "Wrong argument");
+        d[kStart] = val;
+        UnrollVarArgsAssignImpl<T, kStart + 1, kEnd, kStart + 1 == kEnd>::Run(
+                d, args...);
+    }
+};
+
+template <typename T, size_t kStart, size_t kEnd>
+struct UnrollVarArgsAssignImpl<T, kStart, kEnd, true> {
+    __host__ __device__ inline static void Run(T* d) {}
+};
+
+template <typename T>
+struct UnrollVarArgsAssign {
+    template <typename... Args>
+    __host__ __device__ inline static void Run(T* d, Args... args) {
+        UnrollVarArgsAssignImpl<T, 0, sizeof...(Args), sizeof...(Args) == 0>::Run(
+                d, args...);
+    }
+};
+
+template <typename T, size_t N>
+class Array {
+public:
+    static constexpr size_t kSize = N;
+
+    __host__ __device__ inline Array() {}
+
+    template <typename... Args>
+    __host__ __device__ inline explicit Array(const T& val, Args... args) {
+        static_assert(N == sizeof...(Args) + 1, "Invalid argument");
+        UnrollVarArgsAssign<T>::Run(data_, val, args...);
+    }
+
+    __host__ __device__ inline T& operator[](size_t i) { return *(data_ + i); }
+
+    __host__ __device__ inline const T& operator[](size_t i) const {
+        return *(data_ + i);
+    }
+
+private:
+    template <typename U>
+    __host__ __device__ static inline U* advance(U* ptr, size_t i) {
+        return ptr + i;
+    }
+
+    T data_[N];
+};
+
+// ================================  group_norm forward ===========================
+
+// implementation of groupnorm_forward from
+// https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/phi/kernels/gpu/group_norm_kernel.cu#L115
+
+template <typename T>
+__forceinline__ __device__ T
+CudaShuffleDownSync(T val, int delta, int width = warpSize) {
+    return __shfl_down(val, static_cast<unsigned>(delta), width);
+}
+
+template <>
+__forceinline__ __device__ dt_float16
+CudaShuffleDownSync(dt_float16 val, int delta, int width) {
+    return dt_float16(__shfl_down(val, static_cast<unsigned>(delta), width));
+}
+
+template <>
+__forceinline__ __device__ dt_bfloat16
+CudaShuffleDownSync(dt_bfloat16 val, int delta, int width) {
+    return dt_bfloat16(__shfl_down(val, static_cast<unsigned>(delta), width));
+}
+
+template <typename T, int VecSize>
+struct alignas(sizeof(T) * VecSize) VectorType {
+    T val[VecSize];
+};
+
+template <typename T>
+struct AddFunctor {
+    inline T initial() { return static_cast<T>(0.0f); }
+
+    __device__ __forceinline__ T operator()(const T a, const T b) const {
+        return b + a;
+    }
+};
+
+template <typename T, typename ReduceOp>
+__device__ __forceinline__ T WarpReduce(T val, ReduceOp reducer) {
+    for (int stride = WARP_SIZE / 2; stride > 0; stride >>= 1) {
+        T temp = CudaShuffleDownSync<T>(val, stride);
+        val = reducer(val, temp);
+    }
+    return val;
+}
+
+template <typename T, typename ReduceOp>
+__device__ __forceinline__ T BlockXReduce(T val, ReduceOp reducer) {
+    __syncthreads();
+    __shared__ T shared[64];
+    int block_dim_x = blockDim.x;
+    if (blockDim.x > WARP_SIZE) {
+        block_dim_x = blockDim.x / WARP_SIZE;
+        int lane = threadIdx.x % WARP_SIZE;
+        int tid = threadIdx.y * blockDim.x + threadIdx.x;
+        int wid = tid / WARP_SIZE;
+        int bid = threadIdx.y;
+        val = WarpReduce<T, ReduceOp>(val, reducer);
+        if (lane == 0) {
+            shared[wid] = val;
+        }
+        __syncthreads();
+        val = shared[bid * block_dim_x + lane];
+    }
+
+    for (int stride = 1; stride < block_dim_x; stride <<= 1) {
+        T temp = CudaShuffleDownSync(val, stride);
+        val = reducer(val, temp);
+    }
+    if (threadIdx.x == 0) {
+        shared[threadIdx.y] = val;
+    }
+    __syncthreads();
+    return shared[threadIdx.y];
+}
+
+template <typename T>
+__device__ __forceinline__ void ReduceMeanAndVar(
+        T* mean, T* var, T x_mean, T x_var, int size) {
+    const int nc = blockIdx.x;
+    x_mean = BlockXReduce<T, AddFunctor<T>>(x_mean, AddFunctor<T>());
+    x_var = BlockXReduce<T, AddFunctor<T>>(x_var, AddFunctor<T>());
+    __syncthreads();
+    if (threadIdx.x == 0) {
+        mean[nc] = static_cast<T>(x_mean / size);
+        var[nc] = static_cast<T>(x_var / size);
+    }
+}
+
+template <typename T, typename T_ACC, int VecSize, int Num>
+__device__ __forceinline__ void ThreadReduce(
+        Array<const T*, Num> arrs, int size, const int offset, T_ACC* out_mean,
+        T_ACC* out_var) {
+    const T* x = arrs[0];
+    const T* y;
+    if (Num == 2) {
+        y = arrs[1];
+    }
+    using VecT = VectorType<T, VecSize>;
+    int tid = threadIdx.x;
+    if (offset > 0) {
+        x -= offset;
+        if (Num == 2) {
+            y -= offset;
+        }
+        size += offset;
+        if (tid >= offset) {
+            if (Num == 1) {
+                *out_mean += x[tid];
+                *out_var += x[tid] * x[tid];
+            } else if (Num == 2) {
+                *out_mean += y[tid];
+                *out_var += y[tid] * x[tid];
+            }
+        }
+        size -= blockDim.x;
+        x += blockDim.x;
+        if (Num == 2) {
+            y += blockDim.x;
+        }
+    }
+    int remain = size % (VecSize * blockDim.x);
+
+    T ins_x[VecSize];
+    T ins_y[VecSize];
+    VecT* ins_vec_x = reinterpret_cast<VecT*>(&ins_x);
+    VecT* ins_vec_y = reinterpret_cast<VecT*>(&ins_y);
+
+    // vector part
+    for (; VecSize * tid < (size - remain); tid += blockDim.x) {
+        *ins_vec_x = reinterpret_cast<const VecT*>(x)[tid];
+        if (Num == 2) {
+            *ins_vec_y = reinterpret_cast<const VecT*>(y)[tid];
+        }
+
+#pragma unroll
+        for (int i = 0; i < VecSize; ++i) {
+            if (Num == 1) {
+                *out_mean += ins_x[i];
+                *out_var += ins_x[i] * ins_x[i];
+            } else if (Num == 2) {
+                *out_mean += ins_y[i];
+                *out_var += ins_y[i] * ins_x[i];
+            }
+        }
+    }
+
+    // scalar part
+    tid = size - remain + threadIdx.x;
+    for (; tid < size; tid += blockDim.x) {
+        if (Num == 1) {
+            *out_mean += x[tid];
+            *out_var += x[tid] * x[tid];
+        } else if (Num == 2) {
+            *out_mean += y[tid];
+            *out_var += y[tid] * x[tid];
+        }
+    }
+}
+
+template <typename T, typename T_ACC>
+__global__ void ScalarGetMeanAndVar(const T* x, T_ACC* mean, T_ACC* var, int size) {
+    int i = blockIdx.x;
+    T_ACC x_mean = static_cast<T_ACC>(0);
+    T_ACC x_var = static_cast<T_ACC>(0);
+    for (int j = threadIdx.x; j < size; j += blockDim.x) {
+        T val;
+        val = x[i * size + j];
+        x_mean += val;
+        x_var += val * val;
+    }
+    ReduceMeanAndVar<T_ACC>(mean, var, x_mean, x_var, size);
+}
+
+template <typename T, typename T_ACC, int VecSize>
+__global__ void VectorizedGetMeanAndVar(const T* x, T_ACC* mean, T_ACC* var, int size) {
+    int i = blockIdx.x;
+    T_ACC x_mean = static_cast<T_ACC>(0);
+    T_ACC x_var = static_cast<T_ACC>(0);
+    x += i * size;
+    const int input_offset = ((uint64_t)x) % 16 / sizeof(T);
+    Array<const T*, 1> ins;
+    ins[0] = x;
+    ThreadReduce<T, T_ACC, VecSize, 1>(ins, size, input_offset, &x_mean, &x_var);
+    ReduceMeanAndVar<T_ACC>(mean, var, x_mean, x_var, size);
+}
+
+template <typename T, typename T_ACC>
+__global__ void GroupNormForward(
+        const T* x, const T_ACC* mean, const T_ACC* var, const T* scale, const T* bias,
+        int N, int C, int W, int imsize, int groups, int group_size, float epsilon,
+        T* y, T_ACC* real_var) {
+    int gid = blockIdx.y;
+    int cid = blockIdx.x;
+    int bid = blockIdx.z;
+    int ccid = gid * group_size + cid;
+    if (ccid >= C)
+        return;
+    auto ng = bid * groups + gid;
+    T_ACC x_mean = mean[ng];
+    T_ACC x_var = var[ng];
+    x_var = x_var - x_mean * x_mean;
+    T_ACC var_inv = rsqrt(x_var + epsilon);
+    if (cid == 0 && threadIdx.x == 0) {
+        real_var[ng] = x_var;
+    }
+    for (int imid = threadIdx.x; imid < imsize; imid += blockDim.x) {
+        T val;
+        int index = (bid * C + ccid) * imsize + imid;
+        val = x[index];
+        val = (val - x_mean) * var_inv;
+        if (scale != nullptr) {
+            val *= scale[ccid];
+        }
+        if (bias != nullptr) {
+            val += bias[ccid];
+        }
+        y[index] = val;
+    }
+}
+
+template <typename T, typename T_ACC>
+void forward(
+        T* src, T* weight, T* bias, T* dst, T_ACC* mean, T_ACC* rstd, T_ACC* temp_rstd,
+        T_ACC eps, int group, int N, int C, int W, int imsize, cudaStream_t stream) {
+    auto group_size = C / group;
+    int block_size = std::min(1024, imsize);
+    dim3 grid(group_size, group, N);
+    dim3 threads(block_size, 1, 1);
+    int size = group_size * imsize;
+    constexpr int vec_size = sizeof(float4) / sizeof(T);
+    int max_block_size = std::min(size / vec_size, 1024);
+    int block_size_temp = 1;
+    while (block_size_temp < max_block_size) {
+        block_size_temp *= 2;
+    }
+    block_size_temp = std::max(block_size_temp, WARP_SIZE);
+    dim3 grids(N * group);
+    dim3 blocks(block_size_temp);
+    if (size < vec_size * block_size_temp) {
+        ScalarGetMeanAndVar<T, T_ACC>
+                <<<grids, blocks, 0, stream>>>(src, mean, temp_rstd, size);
+        after_kernel_launch();
+    } else {
+        VectorizedGetMeanAndVar<T, T_ACC, vec_size>
+                <<<grids, blocks, 0, stream>>>(src, mean, temp_rstd, size);
+        after_kernel_launch();
+    }
+    GroupNormForward<T, T_ACC><<<grid, threads, 0, stream>>>(
+            src, mean, temp_rstd, weight, bias, N, C, W, imsize, group, group_size, eps,
+            dst, rstd);
+    after_kernel_launch();
+}
+
+// ================================  group_norm backward ===========================
+
+// implementation of groupnorm_backward from
+// https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/phi/kernels/gpu/group_norm_grad_kernel.cu#L253
+
+template <typename T, typename T_ACC>
+__global__ void GetDsDbCUDAKernel(int imsize, const T* x, const T* dy, T* ds, T* db) {
+    const int nc = blockIdx.x;
+    T ds_sum = static_cast<T>(0);
+    T db_sum = static_cast<T>(0);
+    for (int i = threadIdx.x; i < imsize; i += blockDim.x) {
+        const int index = nc * imsize + i;
+        ds_sum += dy[index] * x[index];
+        db_sum += dy[index];
+    }
+    ReduceMeanAndVar<T>(db, ds, db_sum, ds_sum, 1);
+}
+
+template <typename T, typename T_ACC>
+__global__ void GetBiasGradientCUDAKernel(
+        int N, int C, int group, T_ACC epsilon, const T_ACC* mean, const T_ACC* var,
+        const T* ds, const T* db, T* d_scale, T* d_bias) {
+    const int c = blockIdx.x * blockDim.x + threadIdx.x;
+    if (c < C) {
+        const int G = group;
+        const int D = C / G;
+        T sum1 = static_cast<T>(0);
+        T sum2 = static_cast<T>(0);
+        for (int n = 0; n < N; ++n) {
+            const int nc = n * C + c;
+            const int ng = n * G + c / D;
+            sum1 += (d_scale == nullptr)
+                          ? T(0)
+                          : ((ds[nc] - db[nc] * static_cast<T>(mean[ng])) *
+                             static_cast<T>(rsqrt((float)(var[ng] + epsilon))));
+            sum2 += (d_bias == nullptr) ? T(0) : db[nc];
+        }
+        if (d_scale != nullptr) {
+            d_scale[c] = sum1;
+        }
+        if (d_bias != nullptr) {
+            d_bias[c] = sum2;
+        }
+    }
+}
+
+template <typename T>
+__inline__ MEGDNN_DEVICE T warp_reduce_sum(T val) {
+#pragma unroll
+    for (int offset = (warpSize >> 1); offset > 0; offset >>= 1) {
+        val += __shfl_down(val, offset, warpSize);
+    }
+    return val;
+}
+
+template <typename T>
+__inline__ MEGDNN_DEVICE T BlockReduceSum(T val, T* shared) {
+    const int lid = threadIdx.x % warpSize;
+    const int wid = threadIdx.x / warpSize;
+    val = warp_reduce_sum(val);
+    __syncthreads();
+    if (lid == 0) {
+        shared[wid] = val;
+    }
+    __syncthreads();
+    val = (threadIdx.x < blockDim.x / warpSize) ? shared[lid] : T(0);
+    if (wid == 0) {
+        val = warp_reduce_sum(val);
+    }
+    return val;
+}
+
+template <typename T, typename T_ACC, int BlockDim>
+__global__ void GetBackwardParamsCUDAKernel(
+        int imsize, int groups, int group_size, T_ACC epsilon, const T_ACC* mean,
+        const T_ACC* var, const T* scale, const T* ds, const T* db, T* p1, T* p2,
+        T* p3) {
+    const int n = blockIdx.x;
+    const int g = blockIdx.y;
+    const int ng = n * groups + g;
+    T sum1 = static_cast<T>(0);
+    T sum2 = static_cast<T>(0);
+    T var_inv = static_cast<T>(rsqrt(var[ng] + epsilon));
+    for (int64_t i = threadIdx.x; i < group_size; i += blockDim.x) {
+        const int64_t index = ng * group_size + i;
+        const int64_t c = g * group_size + i;
+        const T scale_v = scale == nullptr ? T(1) : static_cast<T>(scale[c]);
+        sum1 += ds[index] * scale_v;
+        sum2 += db[index] * scale_v;
+        const T scale_c = scale == nullptr ? T(0) : static_cast<T>(scale[c]);
+        p1[index] = scale_c * var_inv;
+    }
+
+    __shared__ T ds_shared[WARP_SIZE];
+    __shared__ T db_shared[WARP_SIZE];
+    sum1 = BlockReduceSum<T>(sum1, ds_shared);
+    sum2 = BlockReduceSum<T>(sum2, db_shared);
+
+    if (threadIdx.x == 0) {
+        const T s = T(1) / static_cast<T>(group_size * imsize);
+        const T x = (sum2 * static_cast<T>(mean[ng]) - sum1) * static_cast<T>(var_inv) *
+                    static_cast<T>(var_inv) * static_cast<T>(var_inv) * s;
+        p2[ng] = x;
+        p3[ng] = -x * static_cast<T>(mean[ng]) - sum2 * static_cast<T>(var_inv) * s;
+    }
+}
+
+template <typename T, typename T_ACC>
+__global__ void GetXGradientCUDAKernel(
+        int imsize, int C, int group_size, int groups, T* p1, T* p2, T* p3, const T* x,
+        const T* dy, T* dx) {
+    int cid = blockIdx.x;
+    int gid = blockIdx.y;
+    int bid = blockIdx.z;
+    int ccid = bid * C + gid * group_size + cid;
+    int ng = bid * groups + gid;
+    int nc = gid * group_size + cid;
+    for (int imid = threadIdx.x; imid < imsize; imid += blockDim.x) {
+        int index = (bid * C + nc) * imsize + imid;
+        dx[index] = p1[ccid] * dy[index] + p2[ng] * x[index] + p3[ng];
+    }
+}
+
+template <typename T, typename T_ACC>
+void backward(
+        const T* dY_data, const T* X_data, const T_ACC* mean_data,
+        const T_ACC* rstd_data, const T* weight_data, T* dX_data, T* dweight_data,
+        T* dbias_data, T_ACC eps, int group, int N, int C, int imsize, T* ds, T* db,
+        T* p1, T* p2, T* p3, cudaStream_t stream) {
+    auto group_size = C / group;
+    int block_size = std::min(1024, imsize);
+    const int block_dims = 1024;
+    dim3 grid(group_size, group, N);
+    dim3 threads(block_size, 1, 1);
+    const int max_num_threads = 1024;
+    int max_block_size = std::min(imsize, max_num_threads);
+    int block_size_temp = 1;
+    while (block_size_temp < max_block_size) {
+        block_size_temp *= 2;
+    }
+    block_size_temp = std::max(block_size_temp, WARP_SIZE);
+    dim3 blocks(block_size_temp);
+    GetDsDbCUDAKernel<T, T_ACC>
+            <<<N * C, blocks, 0, stream>>>(imsize, X_data, dY_data, ds, db);
+    after_kernel_launch();
+    bool flag = weight_data != nullptr ? true : false;
+    if (flag) {
+        const int block = 256;
+        GetBiasGradientCUDAKernel<T, T_ACC>
+                <<<(C + block - 1) / block, block, 0, stream>>>(
+                        N, C, group, eps, mean_data, rstd_data, ds, db, dweight_data,
+                        dbias_data);
+        after_kernel_launch();
+    }
+
+    GetBackwardParamsCUDAKernel<T, T_ACC, block_dims>
+            <<<dim3(N, group), block_dims, 0, stream>>>(
+                    imsize, group, group_size, eps, mean_data, rstd_data, weight_data,
+                    ds, db, p1, p2, p3);
+    after_kernel_launch();
+    GetXGradientCUDAKernel<T, T_ACC><<<grid, threads, 0, stream>>>(
+            imsize, C, group_size, group, p1, p2, p3, X_data, dY_data, dX_data);
+    after_kernel_launch();
+}
+
+#define INST(T, T_ACC)                                                              \
+    template void forward<T, T_ACC>(                                                \
+            T*, T*, T*, T*, T_ACC*, T_ACC*, T_ACC*, T_ACC, int, int, int, int, int, \
+            cudaStream_t);                                                          \
+    template void backward<T, T_ACC>(                                               \
+            const T*, const T*, const T_ACC*, const T_ACC*, const T*, T*, T*, T*,   \
+            T_ACC, int, int, int, int, T*, T*, T*, T*, T*, cudaStream_t);
+
+INST(dt_float32, dt_float32)
+INST(dt_float16, dt_float32)
+INST(dt_bfloat16, dt_float32)
+#undef INST
+
+}  // namespace group_norm
+}  // namespace cuda
+}  // namespace megdnn
+
+// vim: syntax=cpp.doxygen

dnn/src/cuda/group_norm/group_norm_cuda.cuh

@@ -0,0 +1,24 @@
+#pragma once
+#include <cuda_runtime_api.h>
+
+namespace megdnn {
+namespace cuda {
+namespace group_norm {
+
+template <typename T, typename T_ACC>
+void forward(
+        T* X, T* gamma, T* beta, T* Y, T_ACC* mean, T_ACC* rstd, T_ACC* tesmp_rstd,
+        T_ACC eps, int group, int N, int C, int W, int imsize, cudaStream_t stream);
+
+template <typename T, typename T_ACC>
+void backward(
+        const T* dY_data, const T* X_data, const T_ACC* mean_data,
+        const T_ACC* rstd_data, const T* gamma_data, T* dX_data, T* dgamma_data,
+        T* dbeta_data, T_ACC eps, int group, int N, int C, int imsize, T* ds, T* db,
+        T* p1, T* p2, T* p3, cudaStream_t stream);
+
+}  // namespace group_norm
+}  // namespace cuda
+}  // namespace megdnn
+
+// vim: syntax=cpp.doxygen

dnn/src/cuda/group_norm/opr_impl.cpp

@@ -0,0 +1,143 @@
+#include "src/cuda/group_norm/opr_impl.h"
+#include "src/cuda/group_norm/group_norm_cuda.cuh"
+#include "src/cuda/utils.h"
+
+namespace megdnn {
+namespace cuda {
+
+size_t GroupNormForwardImpl::get_workspace_in_bytes(
+        const TensorLayout&, const TensorLayout&, const TensorLayout&,
+        const TensorLayout&, const TensorLayout&, const TensorLayout& rstd) {
+    size_t N = rstd.shape[0];
+    size_t G = rstd.shape[1];
+    return get_workspace_bundle(N, G, rstd.dtype.size()).total_size_in_bytes();
+}
+
+WorkspaceBundle GroupNormForwardImpl::get_workspace_bundle(
+        size_t N, size_t G, size_t dtype_size, void* raw_ptr) {
+    return {raw_ptr, {N * G * dtype_size}, handle()->alignment_requirement()};
+}
+
+void GroupNormForwardImpl::exec(
+        _megdnn_tensor_in data, _megdnn_tensor_in weight, _megdnn_tensor_in bias,
+        _megdnn_tensor_out dst, _megdnn_tensor_out mean, _megdnn_tensor_out rstd,
+        _megdnn_workspace workspace) {
+    check_exec(
+            data.layout, weight.layout, bias.layout, dst.layout, mean.layout,
+            rstd.layout, workspace.size);
+
+    auto p = param();
+    using Format = param::GroupNorm::Format;
+    float eps = p.eps;
+    int group = p.group;
+    bool affine = p.affine;
+    auto layout = data.layout;
+    auto format = p.format;
+    size_t N, C, H, W, imsize;
+    if (data.layout.ndim == 4 && format == Format::NCHW) {
+        N = layout.shape[0];
+        C = layout.shape[1];
+        H = layout.shape[2];
+        W = layout.shape[3];
+        imsize = H * W;
+    } else {
+        megdnn_throw(ssprintf("Unspport groupnorm input"));
+    }
+
+    auto stream = cuda_stream(handle());
+    using namespace ::megdnn::cuda::group_norm;
+    auto wbundle =
+            get_workspace_bundle(N, group, rstd.layout.dtype.size(), workspace.raw_ptr);
+
+#define cb(DType)                                                                  \
+    if (data.layout.dtype == DType()) {                                            \
+        using T = typename DTypeTrait<DType>::ctype;                               \
+        using T_ACC = float;                                                       \
+        T_ACC* temp_rstd = wbundle.get_workspace(0).ptr<T_ACC>();                  \
+        forward<T, T_ACC>(                                                         \
+                data.ptr<T>(), affine ? weight.ptr<T>() : nullptr,                 \
+                affine ? bias.ptr<T>() : nullptr, dst.ptr<T>(), mean.ptr<T_ACC>(), \
+                rstd.ptr<T_ACC>(), temp_rstd, static_cast<T_ACC>(eps),             \
+                static_cast<int>(group), static_cast<int>(N), static_cast<int>(C), \
+                static_cast<int>(W), static_cast<int>(imsize), stream);            \
+        return;                                                                    \
+    }
+    MEGDNN_FOREACH_COMPUTING_DTYPE_FLOAT(cb)
+#undef cb
+    megdnn_throw("bad dtype");
+}
+
+size_t GroupNormBackwardImpl::get_workspace_in_bytes(
+        const TensorLayout&, const TensorLayout& data, const TensorLayout&,
+        const TensorLayout& mean, const TensorLayout&, const TensorLayout&,
+        const TensorLayout&, const TensorLayout&) {
+    size_t N = data.shape[0];
+    size_t C = data.shape[1];
+    size_t G = mean.shape[1];
+    return get_workspace_bundle(N, C, G, data.dtype.size()).total_size_in_bytes();
+}
+
+WorkspaceBundle GroupNormBackwardImpl::get_workspace_bundle(
+        size_t N, size_t C, size_t G, size_t dtype_size, void* raw_ptr) {
+    return {raw_ptr,
+            {N * C * dtype_size, N * C * dtype_size, N * C * dtype_size,
+             N * G * dtype_size, N * G * dtype_size},
+            handle()->alignment_requirement()};
+}
+
+void GroupNormBackwardImpl::exec(
+        _megdnn_tensor_in diff, _megdnn_tensor_in data, _megdnn_tensor_in weight,
+        _megdnn_tensor_in mean, _megdnn_tensor_in rstd, _megdnn_tensor_out ddata,
+        _megdnn_tensor_out dweight, _megdnn_tensor_out dbias,
+        _megdnn_workspace workspace) {
+    check_exec(
+            diff.layout, data.layout, weight.layout, mean.layout, rstd.layout,
+            ddata.layout, dweight.layout, dbias.layout, workspace.size);
+    auto p = param();
+    using Format = param::GroupNorm::Format;
+    bool affine = p.affine;
+    float eps = p.eps;
+    int group = p.group;
+    auto layout = data.layout;
+    auto format = p.format;
+    size_t N, C, H, W, imsize;
+    if (layout.ndim == 4 && format == Format::NCHW) {
+        N = layout.shape[0];
+        C = layout.shape[1];
+        H = layout.shape[2];
+        W = layout.shape[3];
+        imsize = H * W;
+    } else {
+        megdnn_throw(ssprintf("Unspport groupnorm input"));
+    }
+
+    auto stream = cuda_stream(handle());
+    using namespace ::megdnn::cuda::group_norm;
+    auto wbundle = get_workspace_bundle(
+            N, C, group, data.layout.dtype.size(), workspace.raw_ptr);
+#define cb(DType)                                                                   \
+    if (data.layout.dtype == DType()) {                                             \
+        using T = typename DTypeTrait<DType>::ctype;                                \
+        using T_ACC = float;                                                        \
+        T* ds = wbundle.get_workspace(0).ptr<T>();                                  \
+        T* db = wbundle.get_workspace(1).ptr<T>();                                  \
+        T* p1 = wbundle.get_workspace(2).ptr<T>();                                  \
+        T* p2 = wbundle.get_workspace(3).ptr<T>();                                  \
+        T* p3 = wbundle.get_workspace(4).ptr<T>();                                  \
+        backward<T, T_ACC>(                                                         \
+                diff.ptr<T>(), data.ptr<T>(), mean.ptr<T_ACC>(), rstd.ptr<T_ACC>(), \
+                affine ? weight.ptr<T>() : nullptr, ddata.ptr<T>(),                 \
+                affine ? dweight.ptr<T>() : nullptr,                                \
+                affine ? dbias.ptr<T>() : nullptr, static_cast<T_ACC>(eps),         \
+                static_cast<int>(group), static_cast<int>(N), static_cast<int>(C),  \
+                static_cast<int>(imsize), ds, db, p1, p2, p3, stream);              \
+        return;                                                                     \
+    }
+    MEGDNN_FOREACH_COMPUTING_DTYPE_FLOAT(cb)
+#undef cb
+    megdnn_throw("bad dtype");
+}
+
+}  // namespace cuda
+}  // namespace megdnn
+// vim: syntax=cpp.doxygen

dnn/src/cuda/group_norm/opr_impl.h

@@ -0,0 +1,47 @@
+#pragma once
+#include "megdnn/oprs.h"
+#include "src/common/utils.h"
+#include "src/cuda/cudnn_wrapper.h"
+
+namespace megdnn {
+namespace cuda {
+
+class GroupNormForwardImpl final : public GroupNormForward {
+public:
+    using GroupNormForward::GroupNormForward;
+    void exec(
+            _megdnn_tensor_in data, _megdnn_tensor_in weight, _megdnn_tensor_in bias,
+            _megdnn_tensor_out dst, _megdnn_tensor_out mean, _megdnn_tensor_out rstd,
+            _megdnn_workspace workspace) override;
+    size_t get_workspace_in_bytes(
+            const TensorLayout&, const TensorLayout&, const TensorLayout&,
+            const TensorLayout&, const TensorLayout&,
+            const TensorLayout& rstd) override;
+
+private:
+    WorkspaceBundle get_workspace_bundle(
+            size_t N, size_t G, size_t dtype_size, void* raw_ptr = nullptr);
+};
+
+class GroupNormBackwardImpl final : public GroupNormBackward {
+public:
+    using GroupNormBackward::GroupNormBackward;
+    void exec(
+            _megdnn_tensor_in diff, _megdnn_tensor_in data, _megdnn_tensor_in weight,
+            _megdnn_tensor_in mean, _megdnn_tensor_in rstd, _megdnn_tensor_out ddata,
+            _megdnn_tensor_out dweight, _megdnn_tensor_out dbias,
+            _megdnn_workspace workspace) override;
+    size_t get_workspace_in_bytes(
+            const TensorLayout&, const TensorLayout& data, const TensorLayout&,
+            const TensorLayout& mean, const TensorLayout&, const TensorLayout&,
+            const TensorLayout&, const TensorLayout&) override;
+
+private:
+    WorkspaceBundle get_workspace_bundle(
+            size_t N, size_t C, size_t G, size_t dtype_size, void* raw_ptr = nullptr);
+};
+
+}  // namespace cuda
+}  // namespace megdnn
+
+// vim: syntax=cpp.doxygen

dnn/src/cuda/handle_create.cpp

@@ -32,6 +32,7 @@
 #include "src/cuda/flip/opr_impl.h"
 #include "src/cuda/gaussian_blur/opr_impl.h"
 #include "src/cuda/group_local/opr_impl.h"
+#include "src/cuda/group_norm/opr_impl.h"
 #include "src/cuda/images2neibs/opr_impl.h"
 #include "src/cuda/indexing_multi_axis_vec/opr_impl.h"
 #include "src/cuda/indexing_one_hot/opr_impl.h"
@@ -163,6 +164,8 @@ MEGDNN_SPECIALIZE_CREATE_OPERATOR(BNBackward);
 MEGDNN_SPECIALIZE_CREATE_OPERATOR(GroupLocalForward);
 MEGDNN_SPECIALIZE_CREATE_OPERATOR(GroupLocalBackwardData);
 MEGDNN_SPECIALIZE_CREATE_OPERATOR(GroupLocalBackwardFilter);
+MEGDNN_SPECIALIZE_CREATE_OPERATOR(GroupNormForward);
+MEGDNN_SPECIALIZE_CREATE_OPERATOR(GroupNormBackward);
 MEGDNN_SPECIALIZE_CREATE_OPERATOR(Flip);
 MEGDNN_SPECIALIZE_CREATE_OPERATOR(Rotate);
 MEGDNN_SPECIALIZE_CREATE_OPERATOR(ROICopy);

dnn/src/naive/group_norm/opr_impl.cpp

@@ -0,0 +1,206 @@
+#include "src/naive/group_norm/opr_impl.h"
+#include <algorithm>
+#include "src/common/utils.h"
+#include "src/naive/handle.h"
+
+using namespace megdnn;
+using namespace naive;
+
+namespace {
+
+using Param = megdnn::GroupNorm::Param;
+
+template <typename T, typename T_ACC = float>
+void forward(
+        _megdnn_tensor_in data, _megdnn_tensor_in weight, _megdnn_tensor_in bias,
+        _megdnn_tensor_out dst, _megdnn_tensor_out mean, _megdnn_tensor_out rstd,
+        const Param& param) {
+    float eps = param.eps;
+    bool affine = param.affine;
+    size_t N = data.layout.shape[0];
+    size_t C = data.layout.shape[1];
+    size_t HxW = data.layout.shape[2] * data.layout.shape[3];
+    const int64_t G = param.group;
+    size_t D = C / G;
+    size_t inner_size = D * HxW;
+
+    for (size_t i = 0; i < N * G; i++) {
+        T_ACC slice_sum = static_cast<T>(0.0f);
+        for (size_t j = 0; j < inner_size; j++) {
+            auto value = data.ptr<T>()[i * inner_size + j];
+            slice_sum += value;
+        }
+        T_ACC slice_mean = static_cast<T>(slice_sum / inner_size);
+
+        T_ACC slice_var = static_cast<T>(0.0f);
+        for (size_t j = 0; j < inner_size; j++) {
+            slice_var += (data.ptr<T>()[i * inner_size + j] - slice_mean) *
+                         (data.ptr<T>()[i * inner_size + j] - slice_mean);
+        }
+        slice_var = slice_var / inner_size;
+
+        T_ACC slice_std = static_cast<T>(1.0f) / static_cast<T>(sqrt(slice_var + eps));
+        if (affine) {
+            const int64_t g = i % G;
+            for (size_t j = 0; j < D; j++) {
+                const int64_t c = g * D + j;
+                T_ACC s = slice_std * weight.ptr<T>()[c];
+                T_ACC b = -s * slice_mean + bias.ptr<T>()[c];
+                for (size_t k = 0; k < HxW; k++) {
+                    dst.ptr<T>()[(i * D + j) * HxW + k] =
+                            s * data.ptr<T>()[(i * D + j) * HxW + k] + b;
+                }
+            }
+        } else {
+            for (size_t j = 0; j < inner_size; j++) {
+                dst.ptr<T>()[i * inner_size + j] =
+                        (data.ptr<T>()[i * inner_size + j] - slice_mean) / slice_std;
+            }
+        }
+        mean.ptr<T_ACC>()[i] = static_cast<T_ACC>(slice_mean);
+        rstd.ptr<T_ACC>()[i] = static_cast<T_ACC>(slice_var);
+    }
+}
+
+template <typename T, typename T_ACC = float>
+void backward(
+        _megdnn_tensor_in diff, _megdnn_tensor_in data, _megdnn_tensor_in weight,
+        _megdnn_tensor_in mean, _megdnn_tensor_in rstd, _megdnn_tensor_out ddata,
+        _megdnn_tensor_out dweight, _megdnn_tensor_out dbias, const Param& param,
+        WorkspaceBundle wbundle) {
+    bool affine = param.affine;
+    size_t N = data.layout.shape[0];
+    size_t C = data.layout.shape[1];
+    size_t G = param.group;
+    float eps = param.eps;
+    size_t HxW = data.layout.shape[2] * data.layout.shape[3];
+    T* ds = wbundle.get_workspace(0).ptr<T>();
+    T* db = wbundle.get_workspace(1).ptr<T>();
+    T* slice_std = wbundle.get_workspace(2).ptr<T>();
+    for (size_t i = 0; i < N * G; i++) {
+        slice_std[i] =
+                static_cast<T>(1.0f) / static_cast<T>(sqrt(rstd.ptr<T_ACC>()[i] + eps));
+    }
+    for (size_t i = 0; i < N * C; i++) {
+        T ds_data = static_cast<T>(0.0f);
+        T db_data = static_cast<T>(0.0f);
+        for (size_t j = 0; j < HxW; j++) {
+            db_data += diff.ptr<T>()[i * HxW + j];
+            ds_data += data.ptr<T>()[i * HxW + j] * diff.ptr<T>()[i * HxW + j];
+        }
+        ds[i] = ds_data;
+        db[i] = db_data;
+    }
+    size_t D = C / G;
+    const T s = T(1) / static_cast<T>(D * HxW);
+    for (size_t i = 0; i < N * G; i++) {
+        const int64_t g = i % G;
+        T ds_v = static_cast<T>(0.0f);
+        T db_v = static_cast<T>(0.0f);
+        for (size_t j = 0; j < D; j += 1) {
+            auto weight_v = affine ? weight.ptr<T>()[g * D + j] : static_cast<T>(1.0f);
+            ds_v += ds[i * D + j] * weight_v;
+            db_v += db[i * D + j] * weight_v;
+        }
+        auto c2 = (db_v * mean.ptr<T_ACC>()[i] - ds_v) * slice_std[i] * slice_std[i] *
+                  slice_std[i] * s;
+        auto c3 = -c2 * mean.ptr<T_ACC>()[i] - db_v * slice_std[i] * s;
+        for (size_t j = 0; j < D; j++) {
+            const int64_t c = g * D + j;
+            auto weight_v = affine ? weight.ptr<T>()[c] : static_cast<T>(1.0f);
+            auto c1 = slice_std[i] * weight_v;
+            for (size_t k = 0; k < HxW; k++) {
+                ddata.ptr<T>()[(i * D + j) * HxW + k] =
+                        c1 * diff.ptr<T>()[(i * D + j) * HxW + k] +
+                        c2 * data.ptr<T>()[(i * D + j) * HxW + k] + c3;
+            }
+        }
+    }
+    if (affine) {
+        for (size_t i = 0; i < C; ++i) {
+            dweight.ptr<T>()[i] = 0;
+            dbias.ptr<T>()[i] = 0;
+        }
+        for (size_t i = 0; i < N * G; i++) {
+            auto g = i % G;
+            for (size_t j = 0; j < D; j++) {
+                auto c = g * D + j;
+                dweight.ptr<T>()[c] +=
+                        (ds[i * D + j] - db[i * D + j] * mean.ptr<T_ACC>()[i]) *
+                        slice_std[i];
+            }
+        }
+        for (size_t i = 0; i < N; i++) {
+            for (size_t j = 0; j < C; j++) {
+                dbias.ptr<T>()[j] += db[i * C + j];
+            }
+        }
+    }
+}
+
+}  // namespace
+
+namespace megdnn {
+namespace naive {
+
+size_t GroupNormBackwardImpl::get_workspace_in_bytes(
+        const TensorLayout&, const TensorLayout& data, const TensorLayout&,
+        const TensorLayout&, const TensorLayout& rstd, const TensorLayout&,
+        const TensorLayout&, const TensorLayout&) {
+    size_t N = data.shape[0];
+    size_t C = data.shape[1];
+    size_t G = rstd.shape[1];
+    return get_workspace_bundle(N, C, G, data.dtype.size()).total_size_in_bytes();
+}
+
+WorkspaceBundle GroupNormBackwardImpl::get_workspace_bundle(
+        size_t N, size_t C, size_t G, size_t dtype_size, void* raw_ptr) {
+    return {raw_ptr,
+            {N * C * dtype_size, N * C * dtype_size, N * G * dtype_size},
+            handle()->alignment_requirement()};
+}
+
+void GroupNormForwardImpl::exec(
+        _megdnn_tensor_in data, _megdnn_tensor_in weight, _megdnn_tensor_in bias,
+        _megdnn_tensor_out dst, _megdnn_tensor_out mean, _megdnn_tensor_out rstd,
+        _megdnn_workspace workspace) {
+    check_exec(
+            data.layout, weight.layout, bias.layout, dst.layout, mean.layout,
+            rstd.layout, workspace.size);
+#define cb(DType)                                                                \
+    if (data.layout.dtype == DType()) {                                          \
+        MEGDNN_DISPATCH_CPU_KERN_OPR(forward<typename DTypeTrait<DType>::ctype>( \
+                data, weight, bias, dst, mean, rstd, param()));                  \
+        return;                                                                  \
+    }
+    MEGDNN_FOREACH_COMPUTING_DTYPE_FLOAT(cb)
+#undef cb
+    megdnn_throw("bad dtype");
+}
+
+void GroupNormBackwardImpl::exec(
+        _megdnn_tensor_in diff, _megdnn_tensor_in data, _megdnn_tensor_in weight,
+        _megdnn_tensor_in mean, _megdnn_tensor_in rstd, _megdnn_tensor_out ddata,
+        _megdnn_tensor_out dweight, _megdnn_tensor_out dbias,
+        _megdnn_workspace workspace) {
+    check_exec(
+            diff.layout, data.layout, weight.layout, mean.layout, rstd.layout,
+            ddata.layout, dweight.layout, dbias.layout, workspace.size);
+#define cb(DType)                                                                 \
+    if (data.layout.dtype == DType()) {                                           \
+        auto wbundle = get_workspace_bundle(                                      \
+                data.layout.shape[0], data.layout.shape[1], rstd.layout.shape[1], \
+                sizeof(DTypeTrait<DType>::ctype), workspace.raw_ptr);             \
+        MEGDNN_DISPATCH_CPU_KERN_OPR(backward<typename DTypeTrait<DType>::ctype>( \
+                diff, data, weight, mean, rstd, ddata, dweight, dbias, param(),   \
+                wbundle));                                                        \
+        return;                                                                   \
+    }
+    MEGDNN_FOREACH_COMPUTING_DTYPE_FLOAT(cb)
+#undef cb
+    megdnn_throw("bad dtype");
+}
+
+}  // namespace naive
+}  // namespace megdnn
+// vim: syntax=cpp.doxygen

dnn/src/naive/group_norm/opr_impl.h

@@ -0,0 +1,44 @@
+#pragma once
+#include "megdnn/oprs.h"
+#include "src/common/utils.h"
+
+namespace megdnn {
+namespace naive {
+
+class GroupNormForwardImpl final : public GroupNormForward {
+public:
+    using GroupNormForward::GroupNormForward;
+    void exec(
+            _megdnn_tensor_in data, _megdnn_tensor_in weight, _megdnn_tensor_in bias,
+            _megdnn_tensor_out dst, _megdnn_tensor_out mean, _megdnn_tensor_out rstd,
+            _megdnn_workspace workspace) override;
+    size_t get_workspace_in_bytes(
+            const TensorLayout&, const TensorLayout&, const TensorLayout&,
+            const TensorLayout&, const TensorLayout&, const TensorLayout&) override {
+        return 0;
+    }
+};
+
+class GroupNormBackwardImpl final : public GroupNormBackward {
+public:
+    using GroupNormBackward::GroupNormBackward;
+    void exec(
+            _megdnn_tensor_in diff, _megdnn_tensor_in data, _megdnn_tensor_in weight,
+            _megdnn_tensor_in mean, _megdnn_tensor_in rstd, _megdnn_tensor_out ddata,
+            _megdnn_tensor_out dweight, _megdnn_tensor_out dbias,
+            _megdnn_workspace workspace) override;
+
+    size_t get_workspace_in_bytes(
+            const TensorLayout&, const TensorLayout& data, const TensorLayout&,
+            const TensorLayout&, const TensorLayout& rstd, const TensorLayout&,
+            const TensorLayout&, const TensorLayout&) override;
+
+private:
+    WorkspaceBundle get_workspace_bundle(
+            size_t N, size_t C, size_t G, size_t dtype_size, void* raw_ptr = nullptr);
+};
+
+}  // namespace naive
+}  // namespace megdnn
+
+// vim: syntax=cpp.doxygen

dnn/src/naive/handle.cpp

@@ -34,6 +34,7 @@
 #include "src/naive/flip/opr_impl.h"
 #include "src/naive/gaussian_blur/opr_impl.h"
 #include "src/naive/group_local/opr_impl.h"
+#include "src/naive/group_norm/opr_impl.h"
 #include "src/naive/images2neibs/opr_impl.h"
 #include "src/naive/indexing_multi_axis_vec/opr_impl.h"
 #include "src/naive/indexing_one_hot/opr_impl.h"

dnn/test/cuda/group_norm.cpp

@@ -0,0 +1,44 @@
+#include "test/cuda/fixture.h"
+
+#include "test/common/checker.h"
+
+namespace megdnn {
+namespace test {
+
+TEST_F(CUDA, GROUPNORM_FORWARD) {
+    using Param = GroupNormForward::Param;
+    Param param;
+    param.affine = true;
+    param.eps = 1e-6;
+    Checker<GroupNormForward> checker(handle_cuda());
+    checker.set_epsilon(1e-2);
+
+    auto run = [&](DType d) {
+        for (size_t group : {1, 3})
+            for (size_t C : {6, 9}) {
+                param.group = group;
+                checker.set_param(param)
+                        .set_dtype(0, d)
+                        .set_dtype(1, d)
+                        .set_dtype(2, d)
+                        .set_dtype(3, d)
+                        .set_dtype(4, dtype::Float32())
+                        .set_dtype(5, dtype::Float32())
+                        .execs({{2, C, 2, 1},
+                                {C},
+                                {C},
+                                {2, C, 2, 1},
+                                {2, group},
+                                {2, group}});
+            }
+    };
+
+    run(dtype::Float32());
+    run(dtype::Float16());
+    run(dtype::BFloat16());
+}
+
+}  // namespace test
+}  // namespace megdnn
+
+// vim: syntax=cpp.doxygen

dnn/test/naive/group_norm.cpp

@@ -0,0 +1,70 @@
+#include "megdnn/dtype.h"
+#include "megdnn/oprs.h"
+#include "test/common/checker.h"
+#include "test/naive/fixture.h"
+
+namespace megdnn {
+namespace test {
+
+TEST_F(NAIVE, GROUPNORM_FORWARD) {
+    Checker<GroupNorm> checker(handle(), true);
+
+    GroupNorm::Param param;
+    param.affine = true;
+    param.group = 3;
+    checker.set_param(param).exect(
+            Testcase{
+                    TensorValue(
+                            {2, 3, 2, 1}, dtype::Float32(),
+                            {3.3179, 0.109, -0.5855, 0.2566, -1.2897, 1.2683, -2.0587,
+                             0.0711, -0.1169, 0.2509, -0.2393, 0.0876}),  // input
+                    TensorValue({3}, dtype::Float32(), {1., 1., 1.}),     // hx
+                    TensorValue({3}, dtype::Float32(), {0., 0., 0.}),     // cx
+                    {},
+                    {},
+                    {}},
+            Testcase{
+                    {},
+                    {},
+                    {},
+                    TensorValue(
+                            {2, 3, 2, 1}, dtype::Float32(),
+                            {1., -1., -1., 1., -1., 1., -1., 1., -0.9999, 0.9999,
+                             -0.9998, 0.9998}),  // output
+                    TensorValue(
+                            {2, 3}, dtype::Float32(),
+                            {1.7135, -0.1645, -0.0107, -0.9938, 0.067,
+                             -0.0758}),  // mean
+                    TensorValue(
+                            {2, 3}, dtype::Float32(),
+                            {2.5742, 0.1772, 1.6358, 1.1340, 0.0338, 0.0267}),  // var
+            });
+    checker.set_param(param).exect(
+            Testcase{
+                    TensorValue(
+                            {1, 3, 1, 2}, dtype::Float32(),
+                            {-2.4348, -1.7948, 0.5223, 0.0932, -0.2955,
+                             -0.0492}),                                // input
+                    TensorValue({3}, dtype::Float32(), {1., 1., 1.}),  // hx
+                    TensorValue({3}, dtype::Float32(), {0., 0., 0.}),  // cx
+                    {},
+                    {},
+                    {}},
+            Testcase{
+                    {},
+                    {},
+                    {},
+                    TensorValue(
+                            {1, 3, 1, 2}, dtype::Float32(),
+                            {-0.9999, 0.9999, 0.9999, -0.9999, -0.9997,
+                             0.9997}),  // output
+                    TensorValue(
+                            {1, 3}, dtype::Float32(),
+                            {-2.1148, 0.3077, -0.1724}),  // mean
+                    TensorValue(
+                            {1, 3}, dtype::Float32(), {0.1023, 0.0460, 0.0151}),  // var
+            });
+}
+
+}  // namespace test
+}  // namespace megdnn

imperative/python/megengine/functional/nn.py

@@ -60,6 +60,7 @@ __all__ = [
     "dropout",
     "embedding",
     "gelu",
+    "group_norm",
     "hsigmoid",
     "hswish",
     "indexing_one_hot",
@@ -1202,6 +1203,33 @@ def softmax(inp: Tensor, axis: Optional[int] = None) -> Tensor:
         return output
 
 
+def group_norm(
+    inp: Tensor,
+    num_groups: int,
+    affine: bool,
+    weight: Optional[Tensor] = None,
+    bias: Optional[Tensor] = None,
+    eps: float = 1e-5,
+):
+    r"""Applies Group Normalization over a mini-batch of inputs as described in
+    the paper `Group Normalization <https://arxiv.org/abs/1803.08494>`__
+    
+    Args:
+        inp: input tensor.
+        num_groups: number of groups to separate the channels into
+        affine: whether to use weight and bias
+        weight: must not be None when the affine is true
+        bias: must not be None when the affine is true
+        eps: a value added to the denominator for numerical stability. Default: 1e-5
+    """
+    op = builtin.GroupNorm(affine=affine, eps=eps, group=num_groups,)
+    if affine:
+        assert weight is not None and bias is not None
+        return apply(op, inp, weight, bias)[0]
+    else:
+        return apply(op, inp)[0]
+
+
 def layer_norm(
     inp: Tensor,
     normalized_shape: tuple,

imperative/python/megengine/module/normalization.py

@@ -34,21 +34,9 @@ class GroupNorm(Module):
             zeros_(self.bias)
 
     def forward(self, x):
-        N, C, H, W = x.shape
-        format = x.format
-        assert C == self.num_channels
-
-        x = x.reshape(N, self.num_groups, -1)
-        mean = x.mean(axis=2, keepdims=True)
-        var = (x * x).mean(axis=2, keepdims=True) - mean * mean
-
-        x = (x - mean) / F.sqrt(var + self.eps)
-        x = x.reshape(N, C, H, W)
-        if self.affine:
-            x = self.weight.reshape(1, -1, 1, 1) * x + self.bias.reshape(1, -1, 1, 1)
-        # FIXME(czh): remove this after making it a builtin op.
-        if format == "nhwc":
-            x = mge.amp.convert_tensor_format(x, inplace=False)
+        x = F.nn.group_norm(
+            x, self.num_groups, self.affine, self.weight, self.bias, self.eps
+        )
         return x
 
     def _module_info_string(self) -> str:

imperative/python/test/unit/module/test_module.py

@@ -8,12 +8,14 @@ import pytest
 import megengine as mge
 import megengine.functional as F
 from megengine import Parameter, Tensor, tensor
+from megengine.device import get_device_count
 from megengine.module import (
     BatchNorm1d,
     BatchNorm2d,
     Conv1d,
     Conv2d,
     Dropout,
+    GroupNorm,
     Linear,
     MaxPool2d,
     Module,
@@ -698,3 +700,67 @@ def test_module_compatible():
     assert (
         old_attributes == current_attributes
     ), "Add or delete attributes in Module class may break compatibility of pickle serialization"
+
+
+def test_grou_norm():
+    class OriginGroupNormFunc(Module):
+        def __init__(self, num_groups, num_channels, eps=1e-5, affine=True, **kwargs):
+            super().__init__(**kwargs)
+            assert num_channels % num_groups == 0
+            self.num_groups = num_groups
+            self.num_channels = num_channels
+            self.eps = eps
+            self.affine = affine
+            if self.affine:
+                self.weight = Parameter(np.ones(num_channels, dtype=np.float32))
+                self.bias = Parameter(np.zeros(num_channels, dtype=np.float32))
+            else:
+                self.weight = None
+                self.bias = None
+
+        def forward(self, x):
+            N, C, H, W = x.shape
+            x = x.reshape(N, self.num_groups, -1)
+            mean = x.mean(axis=2, keepdims=True)
+            var = (x * x).mean(axis=2, keepdims=True) - mean * mean
+            x = (x - mean) / F.sqrt(var + self.eps)
+            x = x.reshape(N, C, H, W)
+            if self.affine:
+                x = self.weight.reshape(1, -1, 1, 1) * x + self.bias.reshape(
+                    1, -1, 1, 1
+                )
+            return x
+
+    inp = np.random.randn(2, 256, 10, 16).astype("float32")
+    mge_inp = Tensor(inp)
+    mge_m = GroupNorm(32, 256)
+
+    ori_inp = Tensor(inp)
+    ori_m = OriginGroupNormFunc(32, 256)
+
+    targets = np.array(2)
+    mge_gm = mge.autodiff.GradManager().attach(mge_m.parameters())
+    ori_gm = mge.autodiff.GradManager().attach(ori_m.parameters())
+
+    for i in range(2):
+        with mge_gm:
+            mge_output = mge_m(mge_inp)
+            loss = F.loss.square_loss(
+                mge_output.sum(), mge.tensor(targets, dtype=np.float32)
+            )
+            mge_gm.backward(loss)
+
+        with ori_gm:
+            ori_output = ori_m(ori_inp)
+            loss = F.loss.square_loss(
+                ori_output.sum(), mge.tensor(targets, dtype=np.float32)
+            )
+            ori_gm.backward(loss)
+
+        np.testing.assert_allclose(mge_output.numpy(), ori_output.numpy(), atol=1e-05)
+        np.testing.assert_allclose(
+            mge_m.weight.grad.numpy(), ori_m.weight.grad.numpy(), rtol=1e-03
+        )
+        np.testing.assert_allclose(
+            mge_m.bias.grad.numpy(), ori_m.bias.grad.numpy(), rtol=1e-03
+        )

imperative/src/impl/ops/group_norm.cpp

@@ -0,0 +1,97 @@
+#include "megbrain/opr/dnn/group_norm.h"
+#include "megbrain/imperative/ops/autogen.h"
+#include "megbrain/opr/internal/megdnn_opr_wrapper.h"
+
+#include "../blob_manager_impl.h"
+#include "../dnn_op_helper.h"
+#include "../op_trait.h"
+
+namespace mgb::imperative {
+namespace group_norm {
+
+cg::OperatorNodeBase* apply_on_var_node(const OpDef& def, const VarNodeArray& inputs) {
+    auto&& op = static_cast<const GroupNorm&>(def);
+    size_t nr_inp = inputs.size();
+    auto p = op.param();
+    mgb_assert((nr_inp == 3 && p.affine) || (nr_inp == 1 && !p.affine));
+    OperatorNodeConfig config{op.make_name()};
+    if (nr_inp == 3) {
+        return opr::GroupNorm::make(
+                       inputs[0], inputs[1], inputs[2], op.param(), config)[0]
+                .node()
+                ->owner_opr();
+    } else {
+        return opr::GroupNorm::make(inputs[0], op.param(), config)[0]
+                .node()
+                ->owner_opr();
+    }
+}
+
+std::tuple<SmallVector<LogicalTensorDesc>, bool> infer_output_attrs_fallible(
+        const OpDef& def, const SmallVector<LogicalTensorDesc>& inputs) {
+    auto&& group_norm = def.cast_final_safe<GroupNorm>();
+    size_t nr_inp = inputs.size();
+    auto affine = group_norm.affine;
+    mgb_assert(
+            (nr_inp == 3 && affine) || (nr_inp == 1 && !affine),
+            "num of inputs of pooling should be 1 or 3 but you give %zu",
+            inputs.size());
+
+    auto&& inp = inputs[0];
+    auto& inp_cn = inp.comp_node;
+
+    if (inp.layout.ndim == 0) {
+        return {{{TensorLayout{inp.layout.dtype}, inp_cn, {}},
+                 {TensorLayout{dtype::Float32()}, inp_cn, {}},
+                 {TensorLayout{dtype::Float32()}, inp_cn, {}}},
+                false};
+    }
+
+    DnnOprHelper<megdnn::GroupNorm> dnn_opr(group_norm.param());
+    auto&& [oup_layout, mean_layout, rstd_layout] =
+            dnn_opr.deduce_layouts<3>(inp.layout, TensorLayout{}, TensorLayout{});
+    return {{{oup_layout, inp_cn, {}},
+             {mean_layout, inp_cn, {}},
+             {rstd_layout, inp_cn, {}}},
+            true};
+}
+
+SmallVector<TensorPtr> apply_on_physical_tensor(
+        const OpDef& def, const SmallVector<TensorPtr>& inputs,
+        SmallVector<LogicalTensorDesc>& output_descs, const bool& validated) {
+    auto&& op_def = def.cast_final_safe<GroupNorm>();
+    size_t nr_inp = inputs.size();
+    auto p = op_def.param();
+
+    mgb_assert(
+            (nr_inp == 3 && p.affine) || (nr_inp == 1 && !p.affine),
+            "num of inputs of groupnorm should be 1 or 3 but you give %zu",
+            inputs.size());
+
+    auto cn = inputs[0]->comp_node();
+    DnnOprCaller<megdnn::GroupNorm> caller(cn, op_def.param());
+
+    auto&& [oup_layout, mean_layout, rstd_layout] = caller.deduce_layouts<3>(
+            inputs[0]->layout(), TensorLayout{}, TensorLayout{});
+
+    auto out = Tensor::make(oup_layout, cn);
+    auto mean = Tensor::make(mean_layout, cn);
+    auto rstd = Tensor::make(rstd_layout, cn);
+
+    if (p.affine) {
+        caller.exec_with_ws(inputs[0], inputs[1], inputs[2], out, mean, rstd);
+    } else {
+        megdnn::TensorND empty_dnn;
+        caller.exec_with_ws(inputs[0], empty_dnn, empty_dnn, out, mean, rstd);
+    }
+    return {out, mean, rstd};
+}
+
+OP_TRAIT_REG(GroupNorm, GroupNorm)
+        .apply_on_var_node(apply_on_var_node)
+        .infer_output_attrs_fallible(infer_output_attrs_fallible)
+        .apply_on_physical_tensor(apply_on_physical_tensor)
+        .fallback();
+
+}  // namespace group_norm
+}  // namespace mgb::imperative

imperative/tablegen/generated/hash.txt

@@ -1,7 +1,7 @@
-905bdf78e5413b06873be64b4ba55db9  ../../dnn/scripts/opr_param_defs.py
-da03ffe2a15411f902cd88920d3d47ec  ../../src/core/include/megbrain/ir/ops.td
-5756619f37e4dc130e1b049d7706d4eb  generated/opdef.h.inl
-98d1291eed73970ee087f898b6241358  generated/opdef.cpp.inl
-b1a9c7569392942294c2168d40939eb5  generated/opdef.py.inl
-3d88d5358d15a39219957f5257e32f5b  generated/opdef.cpy.inl
+e38b68be4e2aaf3de2f22e3dddbeaac4  ../../dnn/scripts/opr_param_defs.py
+cf864561de125ab559c0035158656682  ../../src/core/include/megbrain/ir/ops.td
+9248d42a9b3e770693306992156f6015  generated/opdef.h.inl
+5c7e7ac49d1338d70ac84ba309e6732b  generated/opdef.cpp.inl
+30b669eec36876a65717e0c68dd76c83  generated/opdef.py.inl
+d10455217f5f01e3d2668e5689068920  generated/opdef.cpy.inl
 71e1462bf4d882e2615c3c632cb671cc  generated/enum_macro.h

imperative/tablegen/generated/opdef.cpp.inl

@@ -3775,6 +3775,110 @@ OP_TRAIT_REG(GroupLocal, GroupLocal)
     .props(GroupLocal_props_impl)
     .make_name(GroupLocal_make_name_impl);
 
+MGB_DYN_TYPE_OBJ_FINAL_IMPL(GroupNorm);
+
+namespace {
+size_t GroupNorm_hash_impl(const OpDef& def_) {
+    auto&& op_ = def_.cast_final_safe<GroupNorm>();
+    static_cast<void>(op_);
+    size_t val = mgb::hash(op_.dyn_typeinfo());
+    val = mgb::hash_pair_combine(val, mgb::hash(op_.affine));
+    val = mgb::hash_pair_combine(val, mgb::hash(op_.eps));
+    val = mgb::hash_pair_combine(val, mgb::hash(op_.group));
+    val = mgb::hash_pair_combine(val, mgb::enumhash()(op_.format));
+    return val;
+}
+bool GroupNorm_is_same_st_impl(const OpDef& lhs_, const OpDef& rhs_) {
+    auto &&a_ = lhs_.cast_final_safe<GroupNorm>(),
+         &&b_ = rhs_.cast_final_safe<GroupNorm>();
+    static_cast<void>(a_);
+    static_cast<void>(b_);
+    if (a_.affine != b_.affine) return false;
+    if (a_.eps != b_.eps) return false;
+    if (a_.group != b_.group) return false;
+    if (a_.format != b_.format) return false;
+    return true;
+}
+std::vector<std::pair<const char*, std::string>> GroupNorm_props_impl(const OpDef& def_) {
+    auto&& op_ = def_.cast_final_safe<GroupNorm>();
+    static_cast<void>(op_);
+    std::vector<std::pair<const char*, std::string>> props_;
+    props_.emplace_back("affine", std::to_string(op_.affine));
+    props_.emplace_back("eps", std::to_string(op_.eps));
+    props_.emplace_back("group", std::to_string(op_.group));
+    switch (op_.format){
+    case GroupNorm::Format::NCHW:
+        props_.emplace_back("format", "NCHW");
+        break;
+    case GroupNorm::Format::NHWC:
+        props_.emplace_back("format", "NHWC");
+        break;
+    case GroupNorm::Format::NHWCD4:
+        props_.emplace_back("format", "NHWCD4");
+        break;
+    case GroupNorm::Format::NCHW4:
+        props_.emplace_back("format", "NCHW4");
+        break;
+    case GroupNorm::Format::NCHW8:
+        props_.emplace_back("format", "NCHW8");
+        break;
+    case GroupNorm::Format::NCHW32:
+        props_.emplace_back("format", "NCHW32");
+        break;
+    case GroupNorm::Format::NCHW88:
+        props_.emplace_back("format", "NCHW88");
+        break;
+    case GroupNorm::Format::NCHW44:
+        props_.emplace_back("format", "NCHW44");
+        break;
+    case GroupNorm::Format::NCHW44_DOT:
+        props_.emplace_back("format", "NCHW44_DOT");
+        break;
+    case GroupNorm::Format::NCHW4_NCHW32:
+        props_.emplace_back("format", "NCHW4_NCHW32");
+        break;
+    case GroupNorm::Format::NCHW32_NCHW4:
+        props_.emplace_back("format", "NCHW32_NCHW4");
+        break;
+    case GroupNorm::Format::NCHW4_NCHW:
+        props_.emplace_back("format", "NCHW4_NCHW");
+        break;
+    case GroupNorm::Format::NHWC_NCHW:
+        props_.emplace_back("format", "NHWC_NCHW");
+        break;
+    case GroupNorm::Format::NHWC_NCHW4_IC_SMALL:
+        props_.emplace_back("format", "NHWC_NCHW4_IC_SMALL");
+        break;
+    case GroupNorm::Format::NCHW_NCHW4_IC_SMALL:
+        props_.emplace_back("format", "NCHW_NCHW4_IC_SMALL");
+        break;
+    case GroupNorm::Format::CHWN4:
+        props_.emplace_back("format", "CHWN4");
+        break;
+    case GroupNorm::Format::NCHW64:
+        props_.emplace_back("format", "NCHW64");
+        break;
+    case GroupNorm::Format::NCHW4_NHWC:
+        props_.emplace_back("format", "NCHW4_NHWC");
+        break;
+    default:
+        props_.emplace_back("format", "INVALID");
+        break;
+    }
+    return props_;
+}
+std::string GroupNorm_make_name_impl(const OpDef& def_) {
+    auto&& op_ = def_.cast_final_safe<GroupNorm>();
+    static_cast<void>(op_);
+    return "GroupNorm";
+}
+} // anonymous namespace
+OP_TRAIT_REG(GroupNorm, GroupNorm)
+    .hash(GroupNorm_hash_impl)
+    .is_same_st(GroupNorm_is_same_st_impl)
+    .props(GroupNorm_props_impl)
+    .make_name(GroupNorm_make_name_impl);
+
 MGB_DYN_TYPE_OBJ_FINAL_IMPL(Identity);
 
 namespace {

imperative/tablegen/generated/opdef.cpy.inl

@@ -10075,6 +10075,158 @@ void _init_py_GroupLocal(py::module m) {
     mgb_assert(PyOp(OpDef)::ctype2pytype.emplace(GroupLocal::typeinfo(), &py_type).second);
 }
 
+void _init_py_GroupNorm_Format(PyTypeObject& py_type) {
+    auto& e_type = EnumWrapper<GroupNorm::Format>::type;
+
+    Py_INCREF(e_type);
+    mgb_assert(PyDict_SetItemString(
+        py_type.tp_dict, "Format", reinterpret_cast<PyObject*>(e_type)) >= 0);
+}
+
+PyOpDefBegin(GroupNorm) // {
+    static PyGetSetDef py_getsetters[];
+    static PyMethodDef tp_methods[];
+    
+    static PyObject* getstate(PyObject* self, PyObject*) {
+        auto& opdef = reinterpret_cast<PyOp(GroupNorm)*>(self)->inst();
+        static_cast<void>(opdef);
+        std::unordered_map<std::string, py::object> state {
+            
+            {"affine", serialization<decltype(opdef.affine)>::dump(opdef.affine)},
+            {"eps", serialization<decltype(opdef.eps)>::dump(opdef.eps)},
+            {"group", serialization<decltype(opdef.group)>::dump(opdef.group)},
+            {"format", serialization<decltype(opdef.format)>::dump(opdef.format)}
+        };
+        return py::cast(state).release().ptr();
+    }
+    static PyObject* setstate(PyObject* self, PyObject* args) {
+        PyObject* dict = PyTuple_GetItem(args, 0);
+        if (!dict) return NULL;
+        auto state = py::cast<std::unordered_map<std::string, py::object>>(dict);
+        auto& opdef = reinterpret_cast<PyOp(GroupNorm)*>(self)->inst();
+        static_cast<void>(opdef);
+        
+        {
+        auto&& iter = state.find("affine");
+        if (iter != state.end()) {
+            opdef.affine = serialization<decltype(opdef.affine)>::load(iter->second);
+        }
+        }
+
+        {
+        auto&& iter = state.find("eps");
+        if (iter != state.end()) {
+            opdef.eps = serialization<decltype(opdef.eps)>::load(iter->second);
+        }
+        }
+
+        {
+        auto&& iter = state.find("group");
+        if (iter != state.end()) {
+            opdef.group = serialization<decltype(opdef.group)>::load(iter->second);
+        }
+        }
+
+        {
+        auto&& iter = state.find("format");
+        if (iter != state.end()) {
+            opdef.format = serialization<decltype(opdef.format)>::load(iter->second);
+        }
+        }
+        Py_RETURN_NONE;
+    }
+    static int py_init(PyObject *self, PyObject *args, PyObject *kwds);
+// };
+PyOpDefEnd(GroupNorm)
+
+int PyOp(GroupNorm)::py_init(PyObject *self, PyObject *args, PyObject *kwds) {
+    static const char* kwlist[] = {"affine", "eps", "group", "format", "scope", NULL};
+    PyObject *affine = NULL, *eps = NULL, *group = NULL, *format = NULL, *scope = NULL;
+    if (!PyArg_ParseTupleAndKeywords(args, kwds, "|OOOOO", const_cast<char**>(kwlist), &affine, &eps, &group, &format, &scope))
+    return -1;
+
+    if (affine) {
+        try {
+            // TODO: remove this guard which is used for pybind11 implicit conversion
+            py::detail::loader_life_support guard{};
+            reinterpret_cast<PyOp(GroupNorm)*>(self)->inst().affine =
+                    py::cast<decltype(GroupNorm::affine)>(py::handle(affine));
+        } CATCH_ALL(-1)
+    }
+
+    if (eps) {
+        try {
+            // TODO: remove this guard which is used for pybind11 implicit conversion
+            py::detail::loader_life_support guard{};
+            reinterpret_cast<PyOp(GroupNorm)*>(self)->inst().eps =
+                    py::cast<decltype(GroupNorm::eps)>(py::handle(eps));
+        } CATCH_ALL(-1)
+    }
+
+    if (group) {
+        try {
+            // TODO: remove this guard which is used for pybind11 implicit conversion
+            py::detail::loader_life_support guard{};
+            reinterpret_cast<PyOp(GroupNorm)*>(self)->inst().group =
+                    py::cast<decltype(GroupNorm::group)>(py::handle(group));
+        } CATCH_ALL(-1)
+    }
+
+    if (format) {
+        try {
+            // TODO: remove this guard which is used for pybind11 implicit conversion
+            py::detail::loader_life_support guard{};
+            reinterpret_cast<PyOp(GroupNorm)*>(self)->inst().format =
+                    py::cast<decltype(GroupNorm::format)>(py::handle(format));
+        } CATCH_ALL(-1)
+    }
+
+    if (scope) {
+        try {
+            reinterpret_cast<PyOp(OpDef)*>(self)->op
+                ->set_scope(py::cast<std::string>(py::handle(scope)));
+        } CATCH_ALL(-1)
+    }
+
+    return 0;
+}
+
+PyGetSetDef PyOp(GroupNorm)::py_getsetters[] = {
+    {const_cast<char*>("affine"), py_get_generic(GroupNorm, affine), py_set_generic(GroupNorm, affine), const_cast<char*>("affine"), NULL},
+    {const_cast<char*>("eps"), py_get_generic(GroupNorm, eps), py_set_generic(GroupNorm, eps), const_cast<char*>("eps"), NULL},
+    {const_cast<char*>("group"), py_get_generic(GroupNorm, group), py_set_generic(GroupNorm, group), const_cast<char*>("group"), NULL},
+    {const_cast<char*>("format"), py_get_generic(GroupNorm, format), py_set_generic(GroupNorm, format), const_cast<char*>("format"), NULL},
+    {NULL}  /* Sentinel */
+};
+
+    PyMethodDef PyOp(GroupNorm)::tp_methods[] = {
+        {const_cast<char*>("__getstate__"), PyOp(GroupNorm)::getstate, METH_NOARGS, "GroupNorm getstate"},
+    {const_cast<char*>("__setstate__"), PyOp(GroupNorm)::setstate, METH_VARARGS, "GroupNorm setstate"},
+        {NULL}  /* Sentinel */
+    };
+    
+void _init_py_GroupNorm(py::module m) {
+    using py_op = PyOp(GroupNorm);
+    auto& py_type = PyOpType(GroupNorm);
+    py_type = {PyVarObject_HEAD_INIT(NULL, 0)};
+    py_type.tp_name = "megengine.core._imperative_rt.ops.GroupNorm";
+    py_type.tp_basicsize = sizeof(PyOp(GroupNorm));
+    py_type.tp_flags = Py_TPFLAGS_DEFAULT | Py_TPFLAGS_BASETYPE;
+    py_type.tp_doc = "GroupNorm";
+    py_type.tp_base = &PyOpType(OpDef);
+    py_type.tp_dealloc = py_dealloc_generic<py_op>;
+    py_type.tp_new = py_new_generic<py_op>;
+    py_type.tp_init = py_op::py_init;
+    py_type.tp_methods = py_op::tp_methods;
+    py_type.tp_getset = py_op::py_getsetters;
+    mgb_assert(PyType_Ready(&py_type) >= 0);
+        _init_py_GroupNorm_Format(py_type);
+
+    PyType_Modified(&py_type);
+    m.add_object("GroupNorm", reinterpret_cast<PyObject*>(&py_type));
+    mgb_assert(PyOp(OpDef)::ctype2pytype.emplace(GroupNorm::typeinfo(), &py_type).second);
+}
+
 PyOpDefBegin(Identity) // {
     static PyGetSetDef py_getsetters[];
     static PyMethodDef tp_methods[];
@@ -19237,6 +19389,7 @@ void _init_py_WarpPerspectiveBackwardMat(py::module m) {
     _init_py_GaussianRNG(m); \
     _init_py_GetVarShape(m); \
     _init_py_GroupLocal(m); \
+    _init_py_GroupNorm(m); \
     _init_py_Identity(m); \
     _init_py_Images2Neibs(m); \
     _init_py_IncrMeshIndexing(m); \

imperative/tablegen/generated/opdef.h.inl

@@ -988,6 +988,23 @@ public:
     }
 };
 
+class GroupNorm : public OpDefImplBase<GroupNorm> {
+    MGB_DYN_TYPE_OBJ_FINAL_DECL;
+
+public:
+    using Format = ::megdnn::param::GroupNorm::Format;
+    bool affine = true;
+    float eps = 1e-5f;
+    uint32_t group = 1;
+    Format format = ::megdnn::param::GroupNorm::Format::NCHW;
+    GroupNorm() = default;
+    GroupNorm(bool affine_, float eps_, uint32_t group_, Format format_, std::string scope_ = {}): affine(affine_), eps(eps_), group(group_), format(format_) { set_scope(scope_); }
+    GroupNorm(::megdnn::param::GroupNorm packed_param_0): affine(packed_param_0.affine), eps(packed_param_0.eps), group(packed_param_0.group), format(packed_param_0.format) {}
+    ::megdnn::param::GroupNorm param() const {
+        return {affine, eps, group, format};
+    }
+};
+
 class Identity : public OpDefImplBase<Identity> {
     MGB_DYN_TYPE_OBJ_FINAL_DECL;
 

imperative/tablegen/generated/opdef.py.inl

@@ -1193,6 +1193,17 @@ GroupLocalInst
     .def_readwrite("format", &GroupLocal::format)
     .def_readwrite("compute_mode", &GroupLocal::compute_mode);
 
+py::class_<GroupNorm, std::shared_ptr<GroupNorm>, OpDef> GroupNormInst(m, "GroupNorm");
+
+GroupNormInst.attr("Format") = AdaptivePoolingInst.attr("Format");
+
+GroupNormInst
+    .def(py::init<bool, float, uint32_t, ::megdnn::param::GroupNorm::Format, std::string>(), py::arg("affine") = true, py::arg("eps") = 1e-5f, py::arg("group") = 1, py::arg("format") = ::megdnn::param::GroupNorm::Format::NCHW, py::arg("scope") = {})
+    .def_readwrite("affine", &GroupNorm::affine)
+    .def_readwrite("eps", &GroupNorm::eps)
+    .def_readwrite("group", &GroupNorm::group)
+    .def_readwrite("format", &GroupNorm::format);
+
 py::class_<Identity, std::shared_ptr<Identity>, OpDef> IdentityInst(m, "Identity");
 
 IdentityInst

src/core/include/megbrain/ir/ops.td

@@ -490,6 +490,8 @@ def LRN: MgbHashableOp<"LRN", [LRNParam]>;
 
 def LayerNorm: MgbHashableOp<"LayerNorm", [LayerNormParam]>;
 
+def GroupNorm: MgbHashableOp<"GroupNorm", [GroupNormParam]>;
+
 def LAMBUpdate: MgbHashableOp<"LAMBUpdate", [LAMBUpdateParam]>;
 
 def RNNCell: MgbHashableOp<"RNNCell", [RNNCellParam]>;

src/opr/impl/dnn/dnn.sereg.h

@@ -1,8 +1,10 @@
+#include "megbrain/opr/basic_arith.h"
 #include "megbrain/opr/dnn/adaptive_pooling.h"
 #include "megbrain/opr/dnn/batch_norm.h"
 #include "megbrain/opr/dnn/convolution.h"
 #include "megbrain/opr/dnn/correlation.h"
 #include "megbrain/opr/dnn/fake_quant.h"
+#include "megbrain/opr/dnn/group_norm.h"
 #include "megbrain/opr/dnn/images2neibs.h"
 #include "megbrain/opr/dnn/layer_norm.h"
 #include "megbrain/opr/dnn/local.h"
@@ -15,6 +17,9 @@
 #include "megbrain/opr/dnn/sliding_window_transpose.h"
 #include "megbrain/opr/dnn/softmax.h"
 #include "megbrain/opr/dnn/tqt.h"
+#include "megbrain/opr/io.h"
+#include "megbrain/opr/tensor_manip.h"
+#include "megbrain/serialization/oss_opr_load_dump.h"
 #include "megbrain/serialization/sereg.h"
 #include "megdnn/opr_param_defs.h"
 #include "megdnn/oprs/nn.h"
@@ -524,6 +529,213 @@ struct OprMaker<opr::LayerNormBackward, 0> {
     }
 };
 
+template <>
+struct OprMaker<opr::GroupNorm, 0> {
+    using Param = opr::GroupNorm::Param;
+    static cg::OperatorNodeBase* make(
+            const Param& param, const cg::VarNodeArray& i, ComputingGraph& graph,
+            const OperatorNodeConfig& config) {
+        MGB_MARK_USED_VAR(graph);
+        if (i.size() == 3) {
+            return opr::GroupNorm::make(i[0], i[1], i[2], param, config)[0]
+                    .node()
+                    ->owner_opr();
+        } else {
+            mgb_assert(i.size() == 1);
+            return opr::GroupNorm::make(i[0], param, config)[0].node()->owner_opr();
+        }
+    }
+};
+
+template <>
+struct OprLoadDumpImplV2<opr::GroupNorm, 0> {
+    using Opr = opr::GroupNorm;
+    using Param = opr::GroupNorm::Param;
+    using ElemwiseParam = opr::Elemwise::Param;
+    using ReduceParam = opr::Reduce::Param;
+    static void dump(OprDumpContext& ctx, const cg::OperatorNodeBase& opr) {
+        ctx.write_param<Param>(opr.cast_final_safe<Opr>().param());
+    }
+
+    static cg::OperatorNodeBase* replace_opr(
+            cg::OperatorNodeBase* opr, const VarNodeArray& inputs) {
+        auto graph = inputs[0]->owner_graph();
+        auto comp_node = inputs[0]->comp_node();
+        // std::unique_ptr<StaticInferManager> m_static_infer_manager;
+        auto opr_param = opr->cast_final_safe<opr::GroupNorm>().param();
+        float eps = opr_param.eps;
+        auto half = DTypeScalar(static_cast<megdnn::dt_float32>(0.5));
+        auto param_eps = DTypeScalar(static_cast<megdnn::dt_float32>(eps));
+        auto half_node = opr::ImmutableTensor::make(*graph, half, {comp_node});
+        auto eps_node = opr::ImmutableTensor::make(*graph, param_eps, {comp_node});
+
+        auto origin_shape = opr::GetVarShape::make(inputs[0]).node();
+
+        TensorShape input_shape =
+                inputs[0]->owner_graph()->static_infer_manager().infer_shape(inputs[0]);
+        size_t N = input_shape[0];
+        size_t inner_size = input_shape[1] * input_shape[2] * input_shape[3];
+        int group = opr_param.group;
+        int size = inner_size / group;
+        HostTensorND hv = HostTensorND(inputs[0]->comp_node(), {3}, dtype::Int32());
+        auto* ptr = hv.ptr<dt_int32>();
+        ptr[0] = N;
+        ptr[1] = group;
+        ptr[2] = size;
+        auto target_shape = opr::ImmutableTensor::make(*graph, hv, {comp_node});
+        auto inp = opr::Reshape::make(inputs[0], target_shape);
+
+        auto mean = opr::Reduce::make(inp, {ReduceParam::Mode::MEAN, 2});
+        auto elemwise1 = opr::Elemwise::make({inp, inp}, {ElemwiseParam::Mode::MUL});
+        auto temp_var = opr::Reduce::make(elemwise1, {ReduceParam::Mode::MEAN, 2});
+        auto elemwise2 = opr::Elemwise::make({mean, mean}, {ElemwiseParam::Mode::MUL});
+        auto var =
+                opr::Elemwise::make({temp_var, elemwise2}, {ElemwiseParam::Mode::SUB});
+        auto add_var = opr::Elemwise::make({var, eps_node}, {ElemwiseParam::Mode::ADD});
+        auto sqrt =
+                opr::Elemwise::make({add_var, half_node}, {ElemwiseParam::Mode::POW});
+        auto div = opr::Elemwise::make({inp, mean}, {ElemwiseParam::Mode::SUB});
+        auto temp_inp =
+                opr::Elemwise::make({div, sqrt}, {ElemwiseParam::Mode::TRUE_DIV});
+        auto res = opr::Reshape::make(temp_inp, origin_shape);
+
+        if (inputs.size() == 3) {
+            auto mul_temp =
+                    opr::Elemwise::make({res, inputs[1]}, {ElemwiseParam::Mode::MUL});
+            auto res = opr::Elemwise::make(
+                    {mul_temp, inputs[2]}, {ElemwiseParam::Mode::ADD});
+            return res.node()->owner_opr();
+        } else {
+            return res.node()->owner_opr();
+        }
+    }
+
+    static cg::OperatorNodeBase* load(
+            OprLoadContext& ctx, const cg::VarNodeArray& inputs,
+            const OperatorNodeConfig& config) {
+        // auto& fbs_ctx = CAST_TO_FBS_V2_CTX(ctx);
+        return OprMaker<opr::GroupNorm, 0>::make(
+                ctx.read_param<Param>(), inputs, ctx.graph(), config);
+    }
+};
+
+// OprMaker in MGB_SEREG_OPR only support unique output opr
+template <>
+struct OprMaker<opr::GroupNormBackward, 0> {
+    using Param = opr::GroupNormBackward::Param;
+    static cg::OperatorNodeBase* make(
+            const Param& param, const cg::VarNodeArray& i, ComputingGraph& graph,
+            const OperatorNodeConfig& config) {
+        MGB_MARK_USED_VAR(graph);
+        if (i.size() == 5) {
+            return opr::GroupNormBackward::make(
+                           i[0], i[1], i[2], i[3], i[4], param, config)[0]
+                    .node()
+                    ->owner_opr();
+        } else {
+            mgb_assert(i.size() == 4);
+            return opr::GroupNormBackward::make(
+                           i[0], i[1], i[2], i[3], param, config)[0]
+                    .node()
+                    ->owner_opr();
+        }
+    }
+};
+
+template <>
+struct OprLoadDumpImplV2<opr::GroupNormBackward, 0> {
+    using Opr = opr::GroupNormBackward;
+    using Param = opr::GroupNormBackward::Param;
+    using ElemwiseParam = opr::Elemwise::Param;
+    using ReduceParam = opr::Reduce::Param;
+    static void dump(OprDumpContext& ctx, const cg::OperatorNodeBase& opr) {
+        ctx.write_param<Param>(opr.cast_final_safe<Opr>().param());
+    }
+
+    static cg::OperatorNodeBase* replace_opr(
+            cg::OperatorNodeBase* opr, const VarNodeArray& inputs) {
+        auto rstd = inputs[4];
+        auto graph = inputs[1]->owner_graph();
+        auto comp_node = inputs[1]->comp_node();
+        auto opr_param = opr->cast_final_safe<opr::GroupNormBackward>().param();
+        float eps = opr_param.eps;
+        auto half = DTypeScalar(static_cast<megdnn::dt_float32>(0.5));
+        auto param_eps = DTypeScalar(static_cast<megdnn::dt_float32>(eps));
+        auto half_node = opr::ImmutableTensor::make(*graph, half, {comp_node});
+        auto eps_node = opr::ImmutableTensor::make(*graph, param_eps, {comp_node});
+        auto const_node =
+                opr::ImmutableTensor::make(*graph, DTypeScalar(1), {comp_node});
+
+        TensorShape input_shape =
+                inputs[1]->owner_graph()->static_infer_manager().infer_shape(inputs[0]);
+        auto origin_shape = opr::GetVarShape::make(inputs[1]).node();
+        size_t N = input_shape[0];
+        size_t C = input_shape[1];
+        size_t inner_size = input_shape[1] * input_shape[2] * input_shape[3];
+        int group = opr_param.group;
+        int size = inner_size / group;
+        HostTensorND hv = HostTensorND(inputs[1]->comp_node(), {3}, dtype::Int32());
+        auto* ptr = hv.ptr<dt_int32>();
+        ptr[0] = N;
+        ptr[1] = group;
+        ptr[2] = size;
+        auto target_shape = opr::ImmutableTensor::make(*graph, hv, {comp_node});
+        auto inp = opr::Reshape::make(inputs[1], target_shape);
+
+        auto temp_rstd =
+                opr::Elemwise::make({rstd, eps_node}, {ElemwiseParam::Mode::ADD});
+        auto sqrt =
+                opr::Elemwise::make({temp_rstd, half_node}, {ElemwiseParam::Mode::POW});
+        auto slice_std = opr::Elemwise::make(
+                {const_node, sqrt}, {ElemwiseParam::Mode::TRUE_DIV});
+        auto sub_mean =
+                opr::Elemwise::make({inp, inputs[3]}, {ElemwiseParam::Mode::SUB});
+        auto x_hat =
+                opr::Elemwise::make({sub_mean, slice_std}, {ElemwiseParam::Mode::MUL});
+        x_hat = opr::Reshape::make(x_hat, origin_shape);
+        auto size_node =
+                opr::ImmutableTensor::make(*graph, DTypeScalar(size), {comp_node});
+        auto temp1 = opr::Elemwise::make(
+                {slice_std, size_node}, {ElemwiseParam::Mode::TRUE_DIV});
+
+        auto dx_hat =
+                opr::Elemwise::make({inputs[0], inputs[2]}, {ElemwiseParam::Mode::MUL});
+        HostTensorND tshape = HostTensorND(inputs[1]->comp_node(), {5}, dtype::Int32());
+        auto* ptr2 = tshape.ptr<dt_int32>();
+        ptr2[0] = N;
+        ptr2[1] = group;
+        ptr2[2] = C / group;
+        ptr2[3] = input_shape[2];
+        ptr2[4] = input_shape[3];
+        target_shape = opr::ImmutableTensor::make(*graph, tshape, {comp_node});
+        x_hat = opr::Reshape::make(x_hat, target_shape);
+        dx_hat = opr::Reshape::make(dx_hat, target_shape);
+        auto temp2 =
+                opr::Elemwise::make({size_node, dx_hat}, {ElemwiseParam::Mode::MUL});
+        ptr2[2] = 1;
+        ptr2[3] = 1;
+        ptr2[4] = 1;
+        target_shape = opr::ImmutableTensor::make(*graph, tshape, {comp_node});
+        auto temp3 = opr::Reduce::make(dx_hat, {ReduceParam::Mode::SUM}, target_shape);
+        auto sum_dx_hat =
+                opr::Reduce::make(temp2, {ReduceParam::Mode::SUM}, target_shape);
+        auto temp4 =
+                opr::Elemwise::make({x_hat, sum_dx_hat}, {ElemwiseParam::Mode::MUL});
+        auto temp5 = opr::Elemwise::make({temp2, temp3}, {ElemwiseParam::Mode::SUB});
+        auto temp6 = opr::Elemwise::make({temp5, temp4}, {ElemwiseParam::Mode::SUB});
+        auto dx_temp = opr::Elemwise::make({temp1, temp6}, {ElemwiseParam::Mode::MUL});
+        auto dx = opr::Reshape::make(dx_temp, origin_shape);
+        return dx.node()->owner_opr();
+    }
+
+    static cg::OperatorNodeBase* load(
+            OprLoadContext& ctx, const cg::VarNodeArray& inputs,
+            const OperatorNodeConfig& config) {
+        return OprMaker<opr::GroupNormBackward, 0>::make(
+                ctx.read_param<Param>(), inputs, ctx.graph(), config);
+    }
+};
+
 template <class MegDNNConv = megdnn::LocalShare>
 struct MakeLocalShareCaller2 {
     template <typename Opr>
@@ -747,6 +959,8 @@ MGB_SEREG_OPR(LSQ, 4);
 MGB_SEREG_OPR(LSQBackward, 5);
 MGB_SEREG_OPR(LayerNorm, 0);
 MGB_SEREG_OPR(LayerNormBackward, 0);
+MGB_SEREG_OPR(GroupNorm, 0);
+MGB_SEREG_OPR(GroupNormBackward, 0);
 MGB_SEREG_OPR(RNNCellForward, 6);
 MGB_SEREG_OPR(LSTMCellForward, 7);
 MGB_SEREG_OPR(RNNForward, 3);
@@ -755,6 +969,14 @@ MGB_SEREG_OPR(LSTMForward, 4);
 MGB_SEREG_OPR(LSTMBackward, 9);
 MGB_SEREG_OPR(Softmax, 1);
 MGB_SEREG_OPR(SoftmaxBackward, 2);
+MGB_SEREG_OPR_V2(
+        GroupNorm, 0,
+        (mgb::serialization::OprLoadDumpImplV2<opr::GroupNorm, 0>::replace_opr),
+        VERSION_2, CURRENT_VERSION);
+MGB_SEREG_OPR_V2(
+        GroupNormBackward, 0,
+        (mgb::serialization::OprLoadDumpImplV2<opr::GroupNormBackward, 0>::replace_opr),
+        VERSION_2, CURRENT_VERSION);
 }  // namespace opr
 
 }  // namespace mgb

src/opr/impl/dnn/group_norm.cpp

@@ -0,0 +1,239 @@
+#include "megbrain/opr/dnn/group_norm.h"
+
+#include "megbrain/graph/grad_impl.h"
+#include "megbrain/opr/internal/out_shape_by_sym_var.h"
+#include "megbrain/opr/utility.h"
+
+#include "../internal/megdnn_opr_wrapper.inl"
+
+using namespace mgb;
+using namespace opr;
+
+/* ==================== GroupNormForward  ==================== */
+MGB_DYN_TYPE_OBJ_FINAL_IMPL(GroupNormForward);
+
+GroupNormForward::GroupNormForward(
+        VarNode* data, VarNode* weight, VarNode* bias, const Param& param,
+        const OperatorNodeConfig& config)
+        : Super{data->owner_graph(), config, "group_norm", {data, weight, bias}} {
+    init_megdnn_opr(*this, param);
+
+    add_input({data, weight, bias});
+    output(0)->dtype(data->dtype());
+    output(1)->dtype(dtype::Float32());
+    output(2)->dtype(dtype::Float32());
+}
+
+GroupNormForward::GroupNormForward(
+        VarNode* data, const Param& param, const OperatorNodeConfig& config)
+        : Super{data->owner_graph(), config, "group_norm", {data}} {
+    init_megdnn_opr(*this, param);
+
+    add_input({data});
+    output(0)->dtype(data->dtype());
+    output(1)->dtype(dtype::Float32());
+    output(2)->dtype(dtype::Float32());
+}
+
+SymbolVarArray GroupNormForward::make(
+        SymbolVar data, SymbolVar weight, SymbolVar bias, const Param& param,
+        const OperatorNodeConfig& config) {
+    auto outs = data.node()
+                        ->owner_graph()
+                        ->insert_opr(std::make_unique<GroupNormForward>(
+                                data.node(), weight.node(), bias.node(), param, config))
+                        ->output();
+    SymbolVarArray ret;
+    for (auto&& out : outs) {
+        ret.emplace_back(out);
+    }
+    return ret;
+}
+
+SymbolVarArray GroupNormForward::make(
+        SymbolVar data, const Param& param, const OperatorNodeConfig& config) {
+    auto outs = data.node()
+                        ->owner_graph()
+                        ->insert_opr(std::make_unique<GroupNormForward>(
+                                data.node(), param, config))
+                        ->output();
+    SymbolVarArray ret;
+    for (auto&& out : outs) {
+        ret.emplace_back(out);
+    }
+    return ret;
+}
+
+void GroupNormForward::get_output_var_shape(
+        const TensorShapeArray& inp_shape, TensorShapeArray& out_shape) const {
+    size_t group = param().group;
+    out_shape[0] = inp_shape[0];
+    size_t N = inp_shape[0].shape[0];
+    TensorShape unnormalized_shape{N, group};
+    out_shape[1] = unnormalized_shape;
+    out_shape[2] = unnormalized_shape;
+}
+
+size_t GroupNormForward::get_workspace_size_bytes(
+        const TensorShapeArray& input_shapes,
+        const TensorShapeArray& output_shapes) const {
+    return intl::MegDNNOprMethInvoker<megdnn::GroupNormForward>::get_workspace_in_bytes(
+            megdnn_opr(), this, input_shapes, output_shapes);
+}
+
+void GroupNormForward::scn_do_execute() {
+    if (param().affine) {
+        megdnn_opr()->exec(
+                input(0)->dev_tensor().as_megdnn(), input(1)->dev_tensor().as_megdnn(),
+                input(2)->dev_tensor().as_megdnn(), output(0)->dev_tensor().as_megdnn(),
+                output(1)->dev_tensor().as_megdnn(),
+                output(2)->dev_tensor().as_megdnn(),
+                intl::get_megdnn_workspace_from_var(output().back()));
+    } else {
+        megdnn_opr()->exec(
+                input(0)->dev_tensor().as_megdnn(), {}, {},
+                output(0)->dev_tensor().as_megdnn(),
+                output(1)->dev_tensor().as_megdnn(),
+                output(2)->dev_tensor().as_megdnn(),
+                intl::get_megdnn_workspace_from_var(output().back()));
+    }
+}
+
+#if MGB_ENABLE_GRAD
+MGB_IMPL_OPR_GRAD(GroupNormForward) {
+    auto p = opr.param();
+    SymbolVarArray grad;
+    VarNodeArray ret;
+    if (p.affine) {
+        mgb_assert(wrt_idx < 3, "wrt_idx %zu is out of range", wrt_idx);
+        grad = GroupNormBackward::make(
+                out_grad[0], opr.input(0), opr.input(1), opr.output(1), opr.output(2),
+                opr.param());
+    } else {
+        mgb_assert(wrt_idx < 1, "wrt_idx %zu is out of range", wrt_idx);
+        grad = GroupNormBackward::make(
+                out_grad[0], opr.input(0), opr.output(1), opr.output(2), opr.param());
+    }
+
+    uint32_t nr_ret = p.affine ? 3 : 1;
+    for (uint32_t i = 0; i < nr_ret; ++i) {
+        ret.push_back(grad[i].node());
+    }
+    return ret;
+}
+#endif
+
+/* ==================== GroupNormBackward ==================== */
+MGB_DYN_TYPE_OBJ_FINAL_IMPL(GroupNormBackward);
+
+GroupNormBackward::GroupNormBackward(
+        VarNode* diff, VarNode* data, VarNode* weight, VarNode* mean, VarNode* rstd,
+        const Param& param, const OperatorNodeConfig& config)
+        : Super({diff->owner_graph(),
+                 config,
+                 "group_norm_backward",
+                 {diff, data, weight, mean, rstd}},
+                0, true) {
+    init_megdnn_opr(*this, param);
+    add_input({diff, data, weight, mean, rstd});
+}
+
+GroupNormBackward::GroupNormBackward(
+        VarNode* diff, VarNode* data, VarNode* mean, VarNode* rstd, const Param& param,
+        const OperatorNodeConfig& config)
+        : Super({diff->owner_graph(),
+                 config,
+                 "group_norm_backward",
+                 {diff, data, mean, rstd}},
+                0, true) {
+    init_megdnn_opr(*this, param);
+    add_input({diff, data, mean, rstd});
+    auto mark_empty_var = [&](VarNode* var) {
+        var->add_flag(VarNode::Flag::ALLOW_EMPTY_SHAPE)
+                .add_flag(VarNode::Flag::VOLATILE_CONTENT);
+    };
+    mark_empty_var(output(1));
+    mark_empty_var(output(2));
+}
+
+SymbolVarArray GroupNormBackward::make(
+        SymbolVar diff, SymbolVar data, SymbolVar weight, SymbolVar mean,
+        SymbolVar rstd, const Param& param, const OperatorNodeConfig& config) {
+    auto outs = diff.node()
+                        ->owner_graph()
+                        ->insert_opr(std::make_unique<GroupNormBackward>(
+                                diff.node(), data.node(), weight.node(), mean.node(),
+                                rstd.node(), param, config))
+                        ->output();
+    SymbolVarArray ret;
+    for (auto&& out : outs) {
+        ret.emplace_back(out);
+    }
+    return ret;
+}
+
+SymbolVarArray GroupNormBackward::make(
+        SymbolVar diff, SymbolVar data, SymbolVar mean, SymbolVar rstd,
+        const Param& param, const OperatorNodeConfig& config) {
+    auto outs = diff.node()
+                        ->owner_graph()
+                        ->insert_opr(std::make_unique<GroupNormBackward>(
+                                diff.node(), data.node(), mean.node(), rstd.node(),
+                                param, config))
+                        ->output();
+    SymbolVarArray ret;
+    for (auto&& out : outs) {
+        ret.emplace_back(out);
+    }
+    return ret;
+}
+
+void GroupNormBackward::init_output_static_infer_desc() {
+    using namespace cg::static_infer;
+    auto&& mgr = owner_graph()->static_infer_manager();
+    mgr.register_shape_infer(output(0), ShapeInferDesc::make_identity(input(1)));
+    if (param().affine) {
+        mgr.register_shape_infer(output(1), ShapeInferDesc::make_identity(input(2)));
+        mgr.register_shape_infer(output(2), ShapeInferDesc::make_identity(input(2)));
+    } else {
+        TensorShape empty;
+        empty.ndim = 0;
+        mgr.register_shape_infer(output(1), ShapeInferDesc::make_const(empty));
+        mgr.register_shape_infer(output(2), ShapeInferDesc::make_const(empty));
+    }
+    this->init_output_static_infer_desc_workspace(
+            intl::AutoAddWorkspaceNeedLimitGetter<megdnn::GroupNormBackward>::val);
+}
+
+void GroupNormBackward::init_output_dtype() {
+    output(0)->dtype(input(1)->dtype());
+    output(1)->dtype(input(2)->dtype());
+    output(2)->dtype(input(2)->dtype());
+}
+
+size_t GroupNormBackward::get_workspace_size_bytes(
+        const TensorShapeArray& input_shapes,
+        const TensorShapeArray& output_shapes) const {
+    return intl::MegDNNOprMethInvoker<megdnn::GroupNormBackward>::
+            get_workspace_in_bytes(megdnn_opr(), this, input_shapes, output_shapes);
+}
+
+void GroupNormBackward::scn_do_execute() {
+    if (param().affine) {
+        megdnn_opr()->exec(
+                input(0)->dev_tensor().as_megdnn(), input(1)->dev_tensor().as_megdnn(),
+                input(2)->dev_tensor().as_megdnn(), input(3)->dev_tensor().as_megdnn(),
+                input(4)->dev_tensor().as_megdnn(), output(0)->dev_tensor().as_megdnn(),
+                output(1)->dev_tensor().as_megdnn(),
+                output(2)->dev_tensor().as_megdnn(),
+                intl::get_megdnn_workspace_from_var(output(3)));
+    } else {
+        megdnn_opr()->exec(
+                input(0)->dev_tensor().as_megdnn(), input(1)->dev_tensor().as_megdnn(),
+                {}, input(2)->dev_tensor().as_megdnn(),
+                input(3)->dev_tensor().as_megdnn(), output(0)->dev_tensor().as_megdnn(),
+                {}, {}, intl::get_megdnn_workspace_from_var(output(3)));
+    }
+}
+
+// vim: syntax=cpp.doxygen foldmethod=marker foldmarker=f{{{,f}}}

src/opr/include/megbrain/opr/dnn/group_norm.h

@@ -0,0 +1,67 @@
+#pragma once
+
+#include "megbrain/opr/internal/megdnn_opr_wrapper.h"
+#include "megdnn/oprs.h"
+
+namespace mgb {
+namespace opr {
+
+MGB_DEFINE_OPR_CLASS_WITH_EXPORT(
+        GroupNormForward, intl::MegDNNOprWrapperFwd<megdnn::GroupNormForward>) // {
+public:
+    MGE_WIN_DECLSPEC_FUC GroupNormForward(
+            VarNode* data, VarNode* weight, VarNode* bias, const Param& param,
+            const OperatorNodeConfig& config);
+    MGE_WIN_DECLSPEC_FUC GroupNormForward(
+            VarNode* data, const Param& param, const OperatorNodeConfig& config);
+
+    MGE_WIN_DECLSPEC_FUC static SymbolVarArray make(
+            SymbolVar data, SymbolVar weight, SymbolVar bias, const Param& param = {},
+            const OperatorNodeConfig& config = {});
+    MGE_WIN_DECLSPEC_FUC static SymbolVarArray make(
+            SymbolVar data, const Param& param = {},
+            const OperatorNodeConfig& config = {});
+
+private:
+    void get_output_var_shape(
+            const TensorShapeArray& inp_shape,
+            TensorShapeArray& out_shape) const override;
+    size_t get_workspace_size_bytes(
+            const TensorShapeArray& input_shapes,
+            const TensorShapeArray& output_shapes) const override;
+    void scn_do_execute() override;
+};
+using GroupNorm = GroupNormForward;
+
+MGB_DEFINE_OPR_CLASS_WITH_EXPORT(
+        GroupNormBackward, intl::MegDNNOprWrapperBwd<megdnn::GroupNormBackward>) // {
+public:
+    MGE_WIN_DECLSPEC_FUC GroupNormBackward(
+            VarNode* diff, VarNode* data, VarNode* weight, VarNode* mean, VarNode* rstd,
+            const Param& param, const OperatorNodeConfig& config);
+
+    MGE_WIN_DECLSPEC_FUC GroupNormBackward(
+            VarNode* diff, VarNode* data, VarNode* mean, VarNode* rstd,
+            const Param& param, const OperatorNodeConfig& config);
+
+    MGE_WIN_DECLSPEC_FUC static SymbolVarArray make(
+            SymbolVar diff, SymbolVar data, SymbolVar weight, SymbolVar mean,
+            SymbolVar rstd, const Param& param = {},
+            const OperatorNodeConfig& config = {});
+    MGE_WIN_DECLSPEC_FUC static SymbolVarArray make(
+            SymbolVar diff, SymbolVar data, SymbolVar mean, SymbolVar rstd,
+            const Param& param = {}, const OperatorNodeConfig& config = {});
+
+private:
+    void init_output_static_infer_desc() override;
+    void init_output_dtype() override;
+    size_t get_workspace_size_bytes(
+            const TensorShapeArray& input_shapes,
+            const TensorShapeArray& output_shapes) const override;
+    void scn_do_execute() override;
+};
+
+}  // namespace opr
+}  // namespace mgb
+
+// vim: syntax=cpp.doxygen foldmethod=marker foldmarker=f{{{,f}}}

src/opr/test/dnn/group_norm.cpp

@@ -0,0 +1,90 @@
+#include "megbrain/opr/dnn/group_norm.h"
+#include "megbrain/comp_node_env.h"
+#include "megbrain/test/autocheck.h"
+#include "megbrain/test/helper.h"
+#include "megbrain/test/megdnn_helper.h"
+
+#include "megdnn/oprs.h"
+
+#include <cmath>
+#include <iomanip>
+#include <random>
+#include <sstream>
+
+using namespace mgb;
+
+namespace {
+using Param = opr::GroupNormForward::Param;
+
+void run_forward(bool is_affine) {
+    using Checker = AutoOprChecker<3, 3>;
+
+    Param param;
+    param.eps = 1e-5;
+    param.affine = is_affine;
+    param.group = 3;
+
+    auto make_graph = [&](const Checker::SymInpArray& inputs) -> Checker::SymOutArray {
+        auto out = opr::GroupNormForward::make(inputs[0], inputs[1], inputs[2], param);
+        return {out[0], out[1], out[2]};
+    };
+
+    auto fwd = [&](Checker::NumOutArray& dest, Checker::NumInpArray inp) {
+        auto opr =
+                MegDNNHandle::get(CompNodeEnv::from_comp_node(CompNode::default_cpu()))
+                        ->create_operator<megdnn::GroupNormForward>();
+        auto inp_shape = inp[0]->shape();
+        auto n_slices = inp_shape[0];
+
+        opr->param() = param;
+
+        dest[0].dtype(dtype::Float32())
+                .comp_node(inp[0]->comp_node())
+                .resize(inp_shape);
+        dest[1].dtype(dtype::Float32())
+                .comp_node(inp[0]->comp_node())
+                .resize({n_slices, param.group});
+        dest[2].dtype(dtype::Float32())
+                .comp_node(inp[0]->comp_node())
+                .resize({n_slices, param.group});
+        std::vector<dt_byte> workspace(opr->get_workspace_in_bytes(
+                inp[0]->layout(), inp[1]->layout(), inp[2]->layout(), dest[0].layout(),
+                dest[1].layout(), dest[2].layout()));
+        opr->exec(
+                inp[0]->as_megdnn(), inp[1]->as_megdnn(), inp[2]->as_megdnn(),
+                dest[0].as_megdnn(), dest[1].as_megdnn(), dest[2].as_megdnn(),
+                {workspace.data(), workspace.size()});
+    };
+
+    auto gen = [&](HostTensorND& src) {
+        HostTensorGenerator<dtype::Float32, RandomDistribution::GAUSSIAN> src_gen(0.f);
+        src = *src_gen(src.shape(), src.comp_node());
+    };
+
+    Checker::RunOptions option;
+    option.numdiff_max_err = 1e-4;
+    Checker checker{make_graph, fwd};
+
+    checker.set_input_generator(0, gen);
+    checker.set_input_generator(1, gen);
+    checker.set_input_generator(2, gen);
+    checker.set_input_allow_grad(0, false);
+    checker.set_input_allow_grad(1, false);
+    checker.set_input_allow_grad(2, false);
+    checker.set_output_allow_grad(0, false);
+    checker.set_output_allow_grad(1, false);
+    checker.set_output_allow_grad(2, false);
+
+    checker.run({TensorShape{2, 6, 2, 1}, TensorShape{6}, TensorShape{6}}, option)
+            .run({TensorShape{2, 6, 2, 1}, TensorShape{6}, TensorShape{6}}, option)
+            .run({TensorShape{2, 6, 2, 1}, TensorShape{6}, TensorShape{6}}, option);
+}
+
+TEST(TestOprDNN, GroupNormForward) {
+    REQUIRE_GPU(1);
+    run_forward(true);
+}
+
+}  // anonymous namespace
+
+// vim: syntax=cpp.doxygen foldmethod=marker foldmarker=f{{{,f}}}

src/serialization/impl/schema.fbs

@@ -123,6 +123,7 @@ union OperatorParam {
     param.LSTM = 89,
     param.Softmax = 90,
     param.Diag = 91,
+    param.GroupNorm = 92,
 }
 
 table Operator {

src/serialization/impl/schema_v2.fbs

@@ -140,6 +140,7 @@ union OperatorParam {
     param.LSTM = 89,
     param.Softmax = 90,
     param.Diag = 91,
+    param.GroupNorm = 92,
 }
 
 table Operator {