!1504 add custom tbe ops for quant aware training

Merge pull request !1504 from wandongdong/master

mindspore/nn/layer/quant.py

@@ -22,11 +22,15 @@ from mindspore.common.parameter import Parameter
 from mindspore.common.initializer import initializer
 from mindspore.common.tensor import Tensor
 from mindspore._checkparam import check_int_positive, check_bool, twice
+from mindspore._checkparam import Validator as validator
 from mindspore.nn.cell import Cell
 from mindspore.nn.layer.activation import get_activation
+import mindspore.context as context
+
 
 __all__ = [
     'FakeQuantWithMinMax',
+    'DepthwiseConv2dBatchNormQuant',
     'Conv2dBatchNormQuant',
     'Conv2dQuant',
     'DenseQuant',
@@ -39,6 +43,169 @@ __all__ = [
 ]
 
 
+class BatchNormFoldCell(Cell):
+    """
+    Batch normalization folded.
+
+    Args:
+        momentum (float): Momentum value should be [0, 1]. Default: 0.1.
+        epsilon (float): A small float number to avoid dividing by 0. 1e-5 if dtype in
+            float32 else 1e-3. Default: 1e-5.
+        freeze_bn (int): Delay in steps at which computation switches from regular batch
+            norm to frozen mean and std. Default: 0.
+
+    Inputs:
+        - **x** (Tensor) - Tensor of shape :math:`(N, C, H, W)`.
+        - **mean** (Tensor) - Tensor of shape :math:`(C,)`.
+        - **variance** (Tensor) - Tensor of shape :math:`(C,)`.
+        - **global_step** (Tensor) - Tensor to record current global step.
+
+    Outputs:
+        Tuple of 4 Tensor, the normalized input and the updated parameters.
+
+        - **batch_mean** (Tensor) - Tensor of shape :math:`(C,)`.
+        - **batch_std** (Tensor) - Tensor of shape :math:`(C,)`.
+        - **running_mean** (Tensor) - Tensor of shape :math:`(C,)`.
+        - **running_std** (Tensor) - Tensor of shape :math:`(C,)`.
+
+    """
+
+    def __init__(self, momentum=0.9, epsilon=1e-5, freeze_bn=0):
+        """init batch norm fold layer"""
+        super(BatchNormFoldCell, self).__init__()
+        self.epsilon = epsilon
+        self.is_gpu = context.get_context('device_target') == "GPU"
+        if self.is_gpu:
+            self.bn_train = P.BatchNormFold(momentum, epsilon, is_training=True, freeze_bn=freeze_bn)
+            self.bn_infer = P.BatchNormFold(momentum, epsilon, is_training=False, freeze_bn=freeze_bn)
+        else:
+            self.bn_reduce = P.BNTrainingReduce()
+            self.bn_update = P.BatchNormFoldD(momentum, epsilon, is_training=True, freeze_bn=freeze_bn)
+
+    def construct(self, x, mean, variance, global_step):
+        if self.is_gpu:
+            if self.training:
+                batch_mean, batch_std, running_mean, running_std = self.bn_train(x, mean, variance, global_step)
+            else:
+                batch_mean, batch_std, running_mean, running_std = self.bn_infer(x, mean, variance, global_step)
+        else:
+            if self.training:
+                x_sum, x_square_sum = self.bn_reduce(x)
+                _, batch_mean, batch_std, running_mean, running_std, mean_updated, variance_updated = \
+                    self.bn_update(x, x_sum, x_square_sum, mean, variance)
+                P.Assign()(mean, mean_updated)
+                P.Assign()(variance, variance_updated)
+            else:
+                batch_mean = P.ZerosLike()(variance)
+                batch_std = P.OnesLike()(variance)
+                running_mean = P.TensorAdd()(mean, 0.)
+                running_std = P.Sqrt()(P.TensorAdd()(variance, self.epsilon))
+        return batch_mean, batch_std, running_mean, running_std
+
+
+class FakeQuantWithMinMaxD(Cell):
+    r"""
+    Aware Quantization training op of ascend. This OP provide Fake quantization observer
+    function on data with min and max.
+
+    Args:
+        min_init (int, list): The dimension of channel or 1(layer). Default: -6.
+        max_init (int, list): The dimension of channel or 1(layer). Default: 6.
+        num_bits (int): Quantization number bit, support 4 and 8bit. Default: 8.
+        ema (bool): Exponential Moving Average algorithm update min and max. Default: False.
+        ema_decay (float): Exponential Moving Average algorithm parameter. Default: 0.9999.
+        per_channel (bool): Quantization by layer or channel. Default: False.
+        out_channels (int): declarate the min and max channel size, Default: 1.
+        quant_delay (int): Quantization delay parameters according by global step. Default: 0.
+        symmetric (bool): Quantization algorithm use symmetric or not. Default: False.
+        narrow_range (bool): Quantization algorithm use narrow range or not. Default: False.
+
+    Inputs:
+        - **x** (Tensor) - The input of FakeQuantWithMinMax.
+
+    Outputs:
+        Tensor, with the same type and shape as the `x`.
+
+    Examples:
+        >>> fake_quant = nn.FakeQuantWithMinMaxD()
+        >>> input_x = Tensor(np.array([[1, 2, 1], [-2, 0, -1]]), mindspore.float32)
+        >>> result = fake_quant(input_x)
+    """
+    def __init__(self,
+                 min_init=-6,
+                 max_init=6,
+                 num_bits=8,
+                 ema=False,
+                 ema_decay=0.999,
+                 per_channel=False,
+                 channel_size=1,
+                 quant_delay=0,
+                 symmetric=False,
+                 narrow_range=False,
+                 training=True):
+        """init FakeQuantWithMinMax ascend layer"""
+        super(FakeQuantWithMinMaxD, self).__init__()
+
+        self.min_init = min_init
+        self.num_bits = num_bits
+        self.max_init = max_init
+        self.ema = ema
+        self.ema_decay = ema_decay
+        self.per_channel = per_channel
+        self.channel_size = channel_size
+        self.quant_delay = quant_delay
+        self.symmetric = symmetric
+        self.narrow_range = narrow_range
+        self.training = training
+
+        if not per_channel:
+            self.fake_quant = P.FakeQuantWithMinMax(num_bits=self.num_bits,
+                                                    ema=self.ema,
+                                                    ema_decay=self.ema_decay,
+                                                    quant_delay=self.quant_delay,
+                                                    symmetric=self.symmetric,
+                                                    narrow_range=self.narrow_range,
+                                                    training=training)
+            self.ema_update = P.FakeQuantWithMinMaxUpdate(num_bits=self.num_bits,
+                                                          ema=self.ema,
+                                                          ema_decay=self.ema_decay,
+                                                          quant_delay=self.quant_delay,
+                                                          symmetric=self.symmetric,
+                                                          narrow_range=self.narrow_range,
+                                                          training=training)
+        else:
+            raise RuntimeError("not support per channel")
+
+        if isinstance(min_init, Parameter):
+            self.minq = min_init
+            self.maxq = max_init
+        else:
+            self.minq = Parameter(Tensor(np.array([min_init]).astype(np.float32)),
+                                  name='quant_min',
+                                  requires_grad=False)
+            self.maxq = Parameter(Tensor(np.array([max_init]).astype(np.float32)),
+                                  name='quant_max',
+                                  requires_grad=False)
+        self.reduce_min = P.ReduceMin()
+        self.reduce_max = P.ReduceMax()
+
+    def extend_repr(self):
+        s = 'min_init={}, max_init={}, ema={}, ema_decay={},  per_channel={}, channel_size={}, quant_delay={}'.format(
+            self.min_init, self.max_init, self.ema, self.ema_decay, self.per_channel, self.channel_size,
+            self.quant_delay)
+        return s
+
+    def construct(self, x, minq, maxq):
+        if self.training:
+            min_up, max_up = self.ema_update(x, minq, maxq)
+            out = self.fake_quant(x, min_up, max_up)
+            P.Assign()(self.minq, min_up)
+            P.Assign()(self.maxq, max_up)
+        else:
+            out = self.fake_quant(x, minq, maxq)
+        return out
+
+
 class FakeQuantWithMinMax(Cell):
     r"""
     Aware Quantization training op. This OP provide Fake quantization observer function on data with min and max.
@@ -62,7 +229,7 @@ class FakeQuantWithMinMax(Cell):
         Tensor, with the same type and shape as the `x`.
 
     Examples:
-        >>> fake_quant = nn.FakeQuantWithMinMax()
+        >>> fake_quant = FakeQuantWithMinMax()
         >>> input_x = Tensor(np.array([[1, 2, 1], [-2, 0, -1]]), mindspore.float32)
         >>> result = fake_quant(input_x)
     """
@@ -77,7 +244,9 @@ class FakeQuantWithMinMax(Cell):
                  out_channels=1,
                  quant_delay=0,
                  symmetric=False,
-                 narrow_range=False):
+                 narrow_range=False,
+                 training=True):
+        """init FakeQuantWithMinMax layer"""
         super(FakeQuantWithMinMax, self).__init__()
 
         self.min_init = min_init
@@ -90,12 +259,13 @@ class FakeQuantWithMinMax(Cell):
         self.quant_delay = quant_delay
         self.symmetric = symmetric
         self.narrow_range = narrow_range
+        self.training = training
 
         if per_channel:
-            min_array = np.array([self.min_init for i in range(
+            min_array = np.array([self.min_init for i in range(0, self.out_channels)]).astype(np.float32)
-                0, self.out_channels)]).astype(np.float32)
+            max_array = np.array([self.max_init for i in range(0, self.channel_size)]).astype(np.float32)
-            max_array = np.array([self.max_init for i in range(
+            self.minq = Parameter(Tensor(min_array), name='quant_min', requires_grad=False)
-                0, self.out_channels)]).astype(np.float32)
+            self.maxq = Parameter(Tensor(max_array), name='quant_max', requires_grad=False)
             self.fake_quant_train = P.FakeQuantWithMinMaxPerChannel(num_bits=self.num_bits,
                                                                     ema=self.ema,
                                                                     ema_decay=self.ema_decay,
@@ -113,25 +283,44 @@ class FakeQuantWithMinMax(Cell):
         else:
             min_array = np.array([min_init]).reshape(1).astype(np.float32)
             max_array = np.array([max_init]).reshape(1).astype(np.float32)
-            self.fake_quant_train = P.FakeQuantWithMinMax(num_bits=self.num_bits,
+            self.minq = Parameter(Tensor(min_array), name='quant_min', requires_grad=False)
-                                                          ema=self.ema,
+            self.maxq = Parameter(Tensor(max_array), name='quant_max', requires_grad=False)
-                                                          ema_decay=self.ema_decay,
+            if context.get_context('device_target') == "Ascend":
-                                                          quant_delay=self.quant_delay,
+                self.fake_quant_train = FakeQuantWithMinMaxD(num_bits=self.num_bits,
-                                                          symmetric=self.symmetric,
+                                                             ema=self.ema,
-                                                          narrow_range=self.narrow_range,
+                                                             ema_decay=self.ema_decay,
-                                                          training=True)
+                                                             quant_delay=self.quant_delay,
-            self.fake_quant_infer = P.FakeQuantWithMinMax(num_bits=self.num_bits,
+                                                             symmetric=self.symmetric,
-                                                          ema=self.ema,
+                                                             narrow_range=self.narrow_range,
-                                                          ema_decay=self.ema_decay,
+                                                             training=True,
-                                                          quant_delay=self.quant_delay,
+                                                             min_init=self.minq,
-                                                          symmetric=self.symmetric,
+                                                             max_init=self.maxq)
-                                                          narrow_range=self.narrow_range,
+                self.fake_quant_infer = FakeQuantWithMinMaxD(num_bits=self.num_bits,
-                                                          training=False)
+                                                             ema=self.ema,
-
+                                                             ema_decay=self.ema_decay,
-        self.minq = Parameter(
+                                                             quant_delay=self.quant_delay,
-            Tensor(min_array), name='quant_min', requires_grad=False)
+                                                             symmetric=self.symmetric,
-        self.maxq = Parameter(
+                                                             narrow_range=self.narrow_range,
-            Tensor(max_array), name='quant_max', requires_grad=False)
+                                                             training=False,
+                                                             min_init=self.minq,
+                                                             max_init=self.maxq)
+            elif context.get_context('device_target') == "GPU":
+                self.fake_quant_train = P.FakeQuantWithMinMax(num_bits=self.num_bits,
+                                                              ema=self.ema,
+                                                              ema_decay=self.ema_decay,
+                                                              quant_delay=self.quant_delay,
+                                                              symmetric=self.symmetric,
+                                                              narrow_range=self.narrow_range,
+                                                              training=True)
+                self.fake_quant_infer = P.FakeQuantWithMinMax(num_bits=self.num_bits,
+                                                              ema=self.ema,
+                                                              ema_decay=ema_decay,
+                                                              quant_delay=quant_delay,
+                                                              symmetric=self.symmetric,
+                                                              narrow_range=self.narrow_range,
+                                                              training=False)
+            else:
+                raise ValueError("Not support platform.")
 
     def extend_repr(self):
         s = 'min={}, max={}, ema={}, ema_decay={}, per_channel={}, quant_delay={}'.format(
@@ -146,6 +335,191 @@ class FakeQuantWithMinMax(Cell):
         return out
 
 
+class DepthwiseConv2dBatchNormQuant(Cell):
+    r"""
+    2D depthwise convolution with BatchNormal op folded layer.
+
+    For a more Detailed overview of Conv2d op.
+
+    Args:
+        in_channels (int): The number of input channel :math:`C_{in}`.
+        out_channels (int): The number of output channel :math:`C_{out}`.
+        kernel_size (Union[int, tuple]): Specifies the height and width of the 2D convolution window.
+        stride (int): Specifies stride for all spatial dimensions with the same value.
+        pad_mode: (str): Specifies padding mode. The optional values are "same", "valid", "pad". Default: "same".
+        padding: (int): Implicit paddings on both sides of the input. Default: 0.
+        eps (int): Parameters for BatchNormal. Default: 1e-5.
+        momentum (int): Parameters for BatchNormal op. Default: 0.9.
+        weight_init (Union[Tensor, str, Initializer, numbers.Number]): Initializer for the
+            convolution kernel. Default: 'None'.
+        beta_init (Union[Tensor, str, Initializer, numbers.Number]): Initializer for the
+            beta vector. Default: 'None'.
+        gamma_init (Union[Tensor, str, Initializer, numbers.Number]): Initializer for the
+            gamma vector. Default: 'None'.
+        mean_init (Union[Tensor, str, Initializer, numbers.Number]): Initializer for the
+            mean vector. Default: 'None'.
+        var_init (Union[Tensor, str, Initializer, numbers.Number]): Initializer for the
+            variance vector. Default: 'None'.
+        quant_delay (int): Quantization delay parameters according by global step. Default: 0.
+        freeze_bn (int): Quantization freeze BatchNormal op according by global step. Default: 100000.
+        fake (bool): Conv2dBatchNormQuant Cell add FakeQuantWithMinMax op or not. Default: True.
+        num_bits (int): Quantization number bit, support 4 and 8bit. Default: 8.
+        per_channel (bool): FakeQuantWithMinMax Parameters. Default: False.
+        symmetric (bool): Quantization algorithm use symmetric or not. Default: False.
+        narrow_range (bool): Quantization algorithm use narrow range or not. Default: False.
+
+    Inputs:
+        - **x** (Tensor) - Tensor of shape :math:`(N, C_{in}, H_{in}, W_{in})`.
+
+    Outputs:
+        Tensor of shape :math:`(N, C_{out}, H_{out}, W_{out})`.
+
+   Examples:
+        >>> quant = nn.DepthwiseConv2dBatchNormQuant(1, 6,
+                                                     kernel_size= (2, 2),
+                                                     stride=(1, 1),
+                                                     pad_mode="valid",
+        >>>                                          dilation=(1, 1))
+        >>> input_x = Tensor(np.random.randint(-2, 2, (2, 1, 1, 3)), mindspore.float32)
+        >>> result = quant(input_x)
+    """
+
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 kernel_size,
+                 stride=1,
+                 pad_mode='same',
+                 padding=0,
+                 dilation=1,
+                 group=1,
+                 eps=1e-5,
+                 momentum=0.997,
+                 weight_init=None,
+                 beta_init=None,
+                 gamma_init=None,
+                 mean_init=None,
+                 var_init=None,
+                 quant_delay=0,
+                 freeze_bn=100000,
+                 fake=True,
+                 num_bits=8,
+                 per_channel=False,
+                 symmetric=False,
+                 narrow_range=False):
+        """init DepthwiseConv2dBatchNormQuant layer"""
+        super(DepthwiseConv2dBatchNormQuant, self).__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.pad_mode = pad_mode
+        self.padding = padding
+        self.dilation = twice(dilation)
+        self.stride = twice(stride)
+        self.group = group
+        self.fake = fake
+        self.freeze_bn = freeze_bn
+        self.momentum = momentum
+        self.quant_delay = quant_delay
+        if isinstance(kernel_size, int):
+            self.kernel_size = (kernel_size, kernel_size)
+        else:
+            self.kernel_size = kernel_size
+        if group > 1:
+            validator.check_integer('group', group, 'in_channels', in_channels, 'Conv2dBatchNormQuant')
+            validator.check_integer('group', group, 'in_channels', out_channels, 'Conv2dBatchNormQuant')
+        self.is_depthwise = group > 1
+
+        channel_multiplier = out_channels // in_channels
+        self.conv = P.DepthwiseConv2dNative(channel_multiplier=channel_multiplier,
+                                            kernel_size=kernel_size,
+                                            stride=stride,
+                                            pad_mode=pad_mode,
+                                            pad=padding)
+
+        if weight_init is None:
+            weight_init = initializer('normal', [channel_multiplier, in_channels, *kernel_size])
+        self.weight = Parameter(weight_init, name='weight')
+        if gamma_init is None:
+            gamma_init = initializer('ones', [out_channels])
+        self.gamma = Parameter(gamma_init, name='gamma')
+        if beta_init is None:
+            beta_init = initializer('zeros', [out_channels])
+        self.beta = Parameter(beta_init, name='beta')
+        if mean_init is None:
+            mean_init = initializer('zeros', [out_channels])
+        self.moving_mean = Parameter(
+            mean_init, name='moving_mean', requires_grad=False)
+        if var_init is None:
+            var_init = initializer('ones', [out_channels])
+        self.moving_variance = Parameter(
+            var_init, name='moving_variance', requires_grad=False)
+
+        self.step = Parameter(initializer(
+            'normal', [1], dtype=mstype.int32), name='step', requires_grad=False)
+
+        self.fake_quant_weight = FakeQuantWithMinMax(min_init=-6,
+                                                     max_init=6,
+                                                     ema=False,
+                                                     num_bits=num_bits,
+                                                     quant_delay=quant_delay,
+                                                     per_channel=per_channel,
+                                                     out_channels=out_channels,
+                                                     symmetric=symmetric,
+                                                     narrow_range=narrow_range)
+        self.batchnorm_fold = BatchNormFoldCell(epsilon=eps, momentum=momentum, freeze_bn=freeze_bn)
+
+        self.correct_mul = P.CorrectionMul(self.is_depthwise)
+        if context.get_context('device_target') == "Ascend":
+            self.batchnorm_fold2_train = P.BatchNormFold2_D(freeze_bn=freeze_bn)
+            self.batchnorm_fold2_infer = P.BatchNormFold2_D(freeze_bn=0)
+        elif context.get_context('device_target') == "GPU":
+            self.batchnorm_fold2_train = P.BatchNormFold2(freeze_bn=freeze_bn)
+            self.batchnorm_fold2_infer = P.BatchNormFold2(freeze_bn=0)
+        else:
+            raise ValueError("Not support platform.")
+        self.one = Tensor(1, mstype.int32)
+        self.assignadd = P.AssignAdd()
+        self.is_gpu = context.get_context('device_target') == "GPU"
+
+    def extend_repr(self):
+        s = 'in_channels={}, out_channels={}, kernel_size={}, stride={}, ' \
+            'pad_mode={}, padding={}, dilation={}, group={}, ' \
+            'fake={}, freeze_bn={}, momentum={}, quant_delay={}'.format(
+                self.in_channels, self.out_channels, self.kernel_size, self.stride,
+                self.pad_mode, self.padding, self.dilation, self.group,
+                self.fake, self.freeze_bn, self.momentum, self.quant_delay)
+        return s
+
+    def construct(self, x):
+        out_conv = self.conv(x, self.weight)
+        # BN fold1
+        batch_mean, batch_std, running_mean, running_std = self.batchnorm_fold(out_conv,
+                                                                               self.moving_mean,
+                                                                               self.moving_variance,
+                                                                               self.step)
+        # fake weight
+        weight = self.correct_mul(self.weight, self.gamma, running_std)
+        if self.fake:
+            weight = self.fake_quant_weight(weight)
+        out = self.conv(x, weight)
+        # BN fold2
+        if self.is_gpu:
+            if self.training:
+                out = self.batchnorm_fold2_train(out, self.beta, self.gamma,
+                                                 batch_std, batch_mean, running_std, running_mean, self.step)
+                F.control_depend(out, self.assignadd(self.step, self.one))
+            else:
+                out = self.batchnorm_fold2_infer(out, self.beta, self.gamma,
+                                                 batch_std, batch_mean, running_std, running_mean, self.step)
+        else:
+            if self.training:
+                out = self.batchnorm_fold2_train(out, self.beta, self.gamma, batch_std, batch_mean, running_std)
+                F.control_depend(out, self.assignadd(self.step, self.one))
+            else:
+                out = self.batchnorm_fold2_infer(out, self.beta, self.gamma, batch_std, batch_mean, running_std)
+        return out
+
+
 class Conv2dBatchNormQuant(Cell):
     r"""
     2D convolution with BatchNormal op folded layer.
@@ -215,6 +589,7 @@ class Conv2dBatchNormQuant(Cell):
                  per_channel=False,
                  symmetric=False,
                  narrow_range=False):
+        """init Conv2dBatchNormQuant layer"""
         super(Conv2dBatchNormQuant, self).__init__()
         self.in_channels = in_channels
         self.out_channels = out_channels
@@ -231,7 +606,6 @@ class Conv2dBatchNormQuant(Cell):
             self.kernel_size = (kernel_size, kernel_size)
         else:
             self.kernel_size = kernel_size
-
         if weight_init is None:
             weight_init = initializer(
                 'normal', [out_channels, in_channels // group, *self.kernel_size])
@@ -254,14 +628,6 @@ class Conv2dBatchNormQuant(Cell):
         self.step = Parameter(initializer(
             'normal', [1], dtype=mstype.int32), name='step', requires_grad=False)
 
-        self.conv = P.Conv2D(out_channel=self.out_channels,
-                             kernel_size=self.kernel_size,
-                             mode=1,
-                             pad_mode=self.pad_mode,
-                             pad=self.padding,
-                             stride=self.stride,
-                             dilation=self.dilation,
-                             group=self.group)
         self.fake_quant_weight = FakeQuantWithMinMax(min_init=-6,
                                                      max_init=6,
                                                      ema=False,
@@ -271,23 +637,29 @@ class Conv2dBatchNormQuant(Cell):
                                                      out_channels=out_channels,
                                                      symmetric=symmetric,
                                                      narrow_range=narrow_range)
-        self.batchnorm_fold_train = P.BatchNormFold(epsilon=eps,
+        self.batchnorm_fold = BatchNormFoldCell(epsilon=eps, momentum=momentum, freeze_bn=freeze_bn)
-                                                    momentum=momentum,
+        self.conv = P.Conv2D(out_channel=out_channels,
-                                                    is_training=True,
+                             kernel_size=kernel_size,
-                                                    freeze_bn=freeze_bn)
+                             mode=1,
-        self.batchnorm_fold_infer = P.BatchNormFold(epsilon=eps,
+                             pad_mode=pad_mode,
-                                                    momentum=momentum,
+                             pad=padding,
-                                                    is_training=False,
+                             stride=stride,
-                                                    freeze_bn=freeze_bn)
+                             dilation=1,
+                             group=group)
         self.correct_mul = P.CorrectionMul()
-        self.relu = P.ReLU()
+        if context.get_context('device_target') == "Ascend":
-        self.batchnorm_fold2 = P.BatchNormFold2(freeze_bn=freeze_bn)
+            self.batchnorm_fold2_train = P.BatchNormFold2_D(freeze_bn=freeze_bn)
-        self.batchnorm_fold2_infer = P.BatchNormFold2(freeze_bn=0)
+            self.batchnorm_fold2_infer = P.BatchNormFold2_D(freeze_bn=0)
+        elif context.get_context('device_target') == "GPU":
+            self.batchnorm_fold2_train = P.BatchNormFold2(freeze_bn=freeze_bn)
+            self.batchnorm_fold2_infer = P.BatchNormFold2(freeze_bn=0)
+        else:
+            raise ValueError("Not support platform.")
         self.one = Tensor(1, mstype.int32)
         self.assignadd = P.AssignAdd()
 
     def extend_repr(self):
-        s = 'input_channels={}, output_channels={}, kernel_size={}, stride={}, ' \
+        s = 'in_channels={}, out_channels={}, kernel_size={}, stride={}, ' \
             'pad_mode={}, padding={}, dilation={}, group={}, ' \
             'fake={}, freeze_bn={}, momentum={}, quant_delay={}'.format(
                 self.in_channels, self.out_channels, self.kernel_size, self.stride,
@@ -296,34 +668,32 @@ class Conv2dBatchNormQuant(Cell):
         return s
 
     def construct(self, x):
-        if self.training:
+        out_conv = self.conv(x, self.weight)
-            beta = self.beta
+        # BN fold1
-            gamma = self.gamma
+        batch_mean, batch_std, running_mean, running_std = self.batchnorm_fold(out_conv,
-            gmean = self.moving_mean
+                                                                               self.moving_mean,
-            gvar = self.moving_variance
+                                                                               self.moving_variance,
-            step = self.step
+                                                                               self.step)
-            out_conv = self.conv(x, self.weight)
+        # fake weight
-            batch_mean, batch_std, running_mean, running_std = self.batchnorm_fold_train(
+        weight = self.correct_mul(self.weight, self.gamma, running_std)
-                out_conv, gmean, gvar, step)
+        if self.fake:
-            # BN fold1
+            weight = self.fake_quant_weight(weight)
-            weight = self.correct_mul(self.weight, gamma, running_std)
+        out = self.conv(x, weight)
-            if self.fake:
+        # BN fold2
-                weight = self.fake_quant_weight(weight)
+        if self.is_gpu:
-            out = self.conv(x, weight)
+            if self.training:
-            # BN fold2
+                out = self.batchnorm_fold2_train(out, self.beta, self.gamma,
-            out = self.batchnorm_fold2(
+                                                 batch_std, batch_mean, running_std, running_mean, self.step)
-                out, beta, gamma, batch_std, batch_mean, running_std, running_mean, step)
+                F.control_depend(out, self.assignadd(self.step, self.one))
-            F.control_depend(out, self.assignadd(self.step, self.one))
+            else:
+                out = self.batchnorm_fold2_infer(out, self.beta, self.gamma,
+                                                 batch_std, batch_mean, running_std, running_mean, self.step)
         else:
-            step = self.step
+            if self.training:
-            batch_mean, batch_std, running_mean, running_std = self.batchnorm_fold_infer(
+                out = self.batchnorm_fold2_train(out, self.beta, self.gamma, batch_std, batch_mean, running_std)
-                x, self.moving_mean, self.moving_variance, step)
+                F.control_depend(out, self.assignadd(self.step, self.one))
-            weight = self.correct_mul(self.weight, self.gamma, running_std)
+            else:
-            if self.fake:
+                out = self.batchnorm_fold2_infer(out, self.beta, self.gamma, batch_std, batch_mean, running_std)
-                weight = self.fake_quant_weight(weight)
-            out = self.conv(x, weight)
-            out = self.batchnorm_fold2_infer(out, self.beta, self.gamma, batch_std, batch_mean,
-                                             running_std, running_mean, step)
         return out
 
 
@@ -434,7 +804,7 @@ class Conv2dQuant(Cell):
         return out
 
     def extend_repr(self):
-        s = 'input_channels={}, output_channels={}, kernel_size={}, stride={}, ' \
+        s = 'in_channels={}, out_channels={}, kernel_size={}, stride={}, ' \
             'pad_mode={}, padding={}, dilation={}, group={}, ' \
             'has_bias={}, quant_delay={}'.format(
                 self.in_channels, self.out_channels, self.kernel_size, self.stride,

mindspore/ops/_grad/grad_quant_ops.py

@@ -22,7 +22,7 @@ from ..composite.multitype_ops.zeros_like_impl import zeros_like
 
 @bprop_getters.register(P.FakeQuantWithMinMax)
 def get_bprop_fakequant_with_minmax(self):
-    """Generate bprop for FakeQuantWithMinMax"""
+    """Generate bprop for FakeQuantWithMinMax for GPU and Ascend"""
     op = P.FakeQuantWithMinMaxGrad(num_bits=self.num_bits, quant_delay=self.quant_delay)
 
     def bprop(x, x_min, x_max, out, dout):
@@ -34,7 +34,7 @@ def get_bprop_fakequant_with_minmax(self):
 
 @bprop_getters.register(P.FakeQuantWithMinMaxPerChannel)
 def get_bprop_fakequant_with_minmax_perchannel(self):
-    """Generate bprop for FakeQuantWithMinMaxPerChannel"""
+    """Generate bprop for FakeQuantWithMinMaxPerChannel for GPU"""
     op = P.FakeQuantWithMinMaxPerChannelGrad(num_bits=self.num_bits, quant_delay=self.quant_delay)
 
     def bprop(x, x_min, x_max, out, dout):
@@ -46,7 +46,7 @@ def get_bprop_fakequant_with_minmax_perchannel(self):
 
 @bprop_getters.register(P.BatchNormFold)
 def get_bprop_batchnorm_fold(self):
-    """Generate bprop for BatchNormFold"""
+    """Generate bprop for BatchNormFold for GPU"""
     op = P.BatchNormFoldGrad(self.epsilon, self.is_training, self.freeze_bn)
 
     def bprop(x, mean, variance, global_step, out, dout):
@@ -58,8 +58,8 @@ def get_bprop_batchnorm_fold(self):
 
 @bprop_getters.register(P.CorrectionMul)
 def get_bprop_correction_mul(self):
-    """Generate bprop for CorrectionMul"""
+    """Generate bprop for CorrectionMul for Ascend and GPU"""
-    grad = P.CorrectionMulGrad()
+    grad = P.CorrectionMulGrad(self.channel_axis)
 
     def bprop(x, batch_std, running_std, out, dout):
         dx, d_batch_std = grad(dout, x, batch_std, running_std)
@@ -70,7 +70,7 @@ def get_bprop_correction_mul(self):
 
 @bprop_getters.register(P.BatchNormFold2)
 def get_bprop_batchnorm_fold2(self):
-    """Generate bprop for CorrectionAdd"""
+    """Generate bprop for BatchNormFold2 for GPU"""
     op_f = P.BatchNormFold2Grad(freeze_bn=self.freeze_bn)
 
     def bprop(x, beta, gamma, batch_std, batch_mean, running_std, running_mean, global_step, out, dout):
@@ -80,3 +80,48 @@ def get_bprop_batchnorm_fold2(self):
                zeros_like(global_step)
 
     return bprop
+
+
+@bprop_getters.register(P.BatchNormFoldD)
+def get_bprop_BatchNormFold(self):
+    """Generate bprop for BatchNormFold for Ascend"""
+    op = P.BatchNormFoldGrad_(self.epsilon, self.is_training, self.freeze_bn)
+
+    def bprop(x, x_sum, x_square_sum, mean, variance, out, dout):
+        dx = op(dout[1], dout[2], x, out[1], out[2])
+        return dx, zeros_like(x_sum), zeros_like(x_square_sum), zeros_like(mean), zeros_like(variance)
+
+    return bprop
+
+
+@bprop_getters.register(P.BNTrainingReduce)
+def get_bprop_BNTrainingReduce(self):
+    def bprop(x, out, dout):
+        return (zeros_like(x),)
+
+    return bprop
+
+
+@bprop_getters.register(P.BatchNormFold2_D)
+def get_bprop_batchnorm_fold2_(self):
+    """Generate bprop for BatchNormFold2 for Ascend"""
+    op_reduce = P.BatchNormFold2GradReduce(freeze_bn=self.freeze_bn)
+    op_f = P.BatchNormFold2GradD(freeze_bn=self.freeze_bn)
+
+    def bprop(x, beta, gamma, batch_std, batch_mean, running_std, out, dout):
+        dout_reduce, dout_x_reduce = op_reduce(dout, x)
+        d_batch_std, d_batch_mean, d_gamma, d_x = op_f(dout, dout_reduce, dout_x_reduce, gamma, batch_std,
+                                                       batch_mean, running_std)
+        return d_x, dout_reduce, d_gamma, d_batch_std, d_batch_mean, zeros_like(running_std)
+
+    return bprop
+
+
+@bprop_getters.register(P.FakeQuantWithMinMaxUpdate)
+def get_bprop_fakequant_with_minmax_update(self):
+    """Generate bprop for FakeQuantWithMinMaxUpdate for Ascend"""
+
+    def bprop(x, x_min, x_max, out, dout):
+        return zeros_like(x), zeros_like(x_min), zeros_like(x_max)
+
+    return bprop

mindspore/ops/_op_impl/_custom_op/batchnorm_fold.py

@@ -0,0 +1,149 @@
+# Copyright 2020 Huawei Technologies Co., Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ============================================================================
+
+"""_BatchNormFold op"""
+
+from mindspore.ops.op_info_register import op_info_register, TBERegOp, DataType
+from te import tvm
+from topi import generic
+from topi.cce import util
+
+batch_norm_op_info = TBERegOp("BatchNormFoldD") \
+    .fusion_type("OPAQUE") \
+    .async_flag(False) \
+    .binfile_name("batchnorm_fold.so") \
+    .compute_cost(10) \
+    .kernel_name("batchnorm_fold") \
+    .partial_flag(True) \
+    .attr("momentum", "optional", "float", "all") \
+    .attr("epsilon", "optional", "float", "all") \
+    .attr("is_training", "optional", "bool", "all") \
+    .attr("freeze_bn", "optional", "int", "all") \
+    .attr("data_format", "optional", "str", "all") \
+    .input(0, "x", False, "required", "all") \
+    .input(1, "x_sum", False, "required", "all") \
+    .input(2, "x_square_sum", False, "required", "all") \
+    .input(3, "mean", False, "required", "all") \
+    .input(4, "variance", False, "required", "all") \
+    .output(0, "y", False, "required", "all") \
+    .output(1, "batch_mean", False, "required", "all") \
+    .output(2, "batch_std", False, "required", "all") \
+    .output(3, "running_mean", False, "required", "all") \
+    .output(4, "running_std", False, "required", "all") \
+    .output(5, "mean_updated", False, "required", "all") \
+    .output(6, "variance_updated", False, "required", "all") \
+    .dtype_format(DataType.F32_5HD, DataType.F32_5HD, DataType.F32_5HD, DataType.F32_5HD, DataType.F32_5HD,
+                  DataType.F32_5HD, DataType.F32_5HD, DataType.F32_5HD, DataType.F32_5HD, DataType.F32_5HD,
+                  DataType.F32_5HD, DataType.F32_5HD) \
+    .get_op_info()
+
+
+@op_info_register(batch_norm_op_info)
+def _batchnorm_fold_tbe():
+    """_BatchNormFold TBE register"""
+    return
+
+
+@util.check_input_type(dict, dict, dict, dict, dict,
+                       dict, dict, dict, dict, dict, dict, dict,
+                       float, float, bool, int, str, str)
+def batchnorm_fold(x, x_sum, x_square_sum, mean, variance,
+                   y, batch_mean, batch_std, running_mean, running_std, mean_updated, variance_updated,
+                   momentum=0.9, epsilon=1e-5, is_training=True, freeze_bn=0, data_format="NCHW",
+                   kernel_name="batchnorm_fold"):
+    """batchnorm_fold TBE op"""
+    momentum = 1.0 - momentum
+    util.check_kernel_name(kernel_name)
+    data_format = data_format.upper()
+    if data_format != "NCHW":
+        raise RuntimeError("The data_format only support NCHW")
+
+    shape_x = x.get("shape")
+    shape_mean = mean.get("shape")
+    shape_variance = variance.get("shape")
+    dtype_x = x.get("dtype")
+    dtype_mean = mean.get("dtype")
+    dtype_variance = variance.get("dtype")
+    for shape in (shape_x, shape_mean, shape_variance):
+        util.check_shape_rule(shape)
+        util.check_tensor_shape_size(shape)
+    check_tuple = ("float16", "float32")
+    for dtype in (dtype_x, dtype_mean, dtype_variance):
+        util.check_dtype_rule(dtype.lower(), check_tuple)
+
+    format_data = x.get("format").upper()
+    if format_data not in ("NCHW", "NC1HWC0"):
+        raise RuntimeError("Format of input only support 4D and 5HD")
+
+    if format_data == "NC1HWC0":
+        if len(shape_x) != 5:
+            raise RuntimeError("batchnorm_fold only support shape 5D"
+                               "when input format is NC1HWC0")
+        shape_mean = (1, shape_x[1], 1, 1, shape_x[4])
+    elif format_data == "NCHW":
+        if len(shape_x) < 2 or len(shape_x) > 4:
+            raise RuntimeError("batchnorm_fold only support shape 2D to 4D")
+        if shape_x[1] != shape_mean[0]:
+            raise RuntimeError("data_format is NCHW, shape_bias must"
+                               "be equal to the second axis of shape_x")
+        shape_mean = (1, shape_x[1],)
+        for _ in range(2, len(shape_x)):
+            shape_mean = shape_mean + (1,)
+
+    x_input = tvm.placeholder(shape_x, name="x_input", dtype=dtype_x.lower())
+    x_sum = tvm.placeholder(shape_mean, name="x_sum", dtype=dtype_x.lower())
+    x_square_sum = tvm.placeholder(shape_mean, name="x_square_sum", dtype=dtype_x.lower())
+    mean = tvm.placeholder(shape_mean, name="mean", dtype=dtype_mean.lower())
+    variance = tvm.placeholder(shape_mean, name="variance", dtype=dtype_variance.lower())
+
+    shape_x = te.lang.cce.util.shape_to_list(x_input.shape)
+    num = shape_x[0] * shape_x[2] * shape_x[3]
+    num_rec = 1.0 / num
+
+    # compute the mean of x
+    batch_mean = te.lang.cce.vmuls(x_sum, num_rec)
+
+    # compute the variance of x
+    variance_div = te.lang.cce.vmuls(x_square_sum, num_rec)
+    mean_square = te.lang.cce.vmul(batch_mean, batch_mean)
+    batch_var_biased = te.lang.cce.vsub(variance_div, mean_square)
+
+    if num == 1:
+        batch_var_scaler = 0.0
+    else:
+        batch_var_scaler = float(num) / (num - 1)
+    batch_variance = te.lang.cce.vmuls(batch_var_biased, batch_var_scaler)
+    batch_std = te.lang.cce.vsqrt(te.lang.cce.vadds(batch_variance, epsilon))
+
+    factor = 1.0 - momentum
+    factor_reverse = momentum
+    mean_mul = te.lang.cce.vmuls(batch_mean, factor)
+    mean_mul_rev = te.lang.cce.vmuls(mean, factor_reverse)
+    mean_updated = te.lang.cce.vadd(mean_mul, mean_mul_rev)
+
+    var_mul = te.lang.cce.vmuls(batch_variance, factor)
+    var_mul_rev = te.lang.cce.vmuls(variance, factor_reverse)
+    variance_updated = te.lang.cce.vadd(var_mul, var_mul_rev)
+
+    y = te.lang.cce.vadds(x_input, 0.0)
+    running_mean = te.lang.cce.vadds(mean, 0.0)
+    running_std = te.lang.cce.vsqrt(te.lang.cce.vadds(variance, epsilon))
+    res = [y, batch_mean, batch_std, running_mean, running_std, mean_updated, variance_updated]
+
+    with tvm.target.cce():
+        sch = generic.auto_schedule(res)
+    config = {"name": kernel_name,
+              "tensor_list": [x_input, x_sum, x_square_sum, mean, variance] + res}
+    te.lang.cce.cce_build_code(sch, config)

mindspore/ops/_op_impl/_custom_op/batchnorm_fold2.py

@@ -0,0 +1,110 @@
+# Copyright 2020 Huawei Technologies Co., Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ============================================================================
+
+"""_BatchNormFold2 op"""
+import te.lang.cce
+from te import tvm
+from te.platform.fusion_manager import fusion_manager
+from topi import generic
+from topi.cce import util
+from mindspore.ops.op_info_register import op_info_register, TBERegOp, DataType
+
+SHAPE_SIZE_LIMIT = 2147483648
+
+batchnorm_fold2_op_info = TBERegOp("BatchNormFold2_D") \
+    .fusion_type("ELEMWISE") \
+    .async_flag(False) \
+    .binfile_name("batchnorm_fold2.so") \
+    .compute_cost(10) \
+    .kernel_name("batchnorm_fold2") \
+    .partial_flag(True) \
+    .op_pattern("formatAgnostic") \
+    .input(0, "x", None, "required", None) \
+    .input(1, "beta", None, "required", None) \
+    .input(2, "gamma", None, "required", None) \
+    .input(3, "batch_std", None, "required", None) \
+    .input(4, "batch_mean", None, "required", None) \
+    .input(5, "running_std", None, "required", None) \
+    .output(0, "y", True, "required", "all") \
+    .dtype_format(DataType.F16_Default, DataType.F16_Default, DataType.F16_Default, DataType.F16_Default,
+                  DataType.F16_Default, DataType.F16_Default, DataType.F16_Default) \
+    .dtype_format(DataType.F16_5HD, DataType.F16_5HD, DataType.F16_5HD, DataType.F16_5HD,
+                  DataType.F16_5HD, DataType.F16_5HD, DataType.F16_5HD) \
+    .dtype_format(DataType.F32_Default, DataType.F32_Default, DataType.F32_Default, DataType.F32_Default,
+                  DataType.F32_Default, DataType.F32_Default, DataType.F32_Default) \
+    .dtype_format(DataType.F32_5HD, DataType.F32_5HD, DataType.F32_5HD, DataType.F32_5HD,
+                  DataType.F32_5HD, DataType.F32_5HD, DataType.F32_5HD) \
+    .get_op_info()
+
+
+@op_info_register(batchnorm_fold2_op_info)
+def _batchnorm_fold2_tbe():
+    """_BatchNormFold2 TBE register"""
+    return
+
+
+@fusion_manager.register("batchnorm_fold2")
+def batchnorm_fold2_compute(x, beta, gamma, batch_std, batch_mean, running_std, kernel_name="batchnorm_fold2"):
+    """_BatchNormFold2 compute"""
+    shape_x = te.lang.cce.util.shape_to_list(x.shape)
+    factor = te.lang.cce.vdiv(running_std, batch_std)
+    factor_b = te.lang.cce.broadcast(factor, shape_x)
+    res = te.lang.cce.vmul(x, factor_b)
+    bias = te.lang.cce.vdiv(batch_mean, batch_std)
+    bias = te.lang.cce.vmul(bias, gamma)
+    bias = te.lang.cce.vsub(beta, bias)
+    bias_b = te.lang.cce.broadcast(bias, shape_x)
+    res = te.lang.cce.vadd(res, bias_b)
+    return res
+
+
+@util.check_input_type(dict, dict, dict, dict, dict, dict, dict, str)
+def batchnorm_fold2(x, beta, gamma, batch_std, batch_mean, running_std, y, kernel_name="batchnorm_fold2"):
+    """_BatchNormFold2 op"""
+    shape = x.get("shape")
+    util.check_kernel_name(kernel_name)
+    util.check_shape_rule(shape)
+    util.check_shape_size(shape, SHAPE_SIZE_LIMIT)
+    check_list = ["float16", "float32"]
+    inp_dtype = x.get("dtype").lower()
+    if not inp_dtype in check_list:
+        raise RuntimeError("Dtype of input only support float16, float32")
+    data_format = x.get("format")
+    ori_format = x.get("ori_format")
+    if data_format.upper() not in ("NC1HWC0", "NCHW"):
+        raise RuntimeError("Un supported data format {}".format(data_format))
+    if data_format.upper() == "NCHW" and ori_format != "NCHW":
+        raise RuntimeError("data_format(NCHW) must same as ori_format")
+    shape_c = gamma.get("shape")
+    if gamma.get("format").upper() == "NCHW":
+        shape_c = 1, gamma.get("shape")[0], 1, 1
+    x_t = tvm.placeholder(shape, name="x", dtype=inp_dtype)
+    beta_t = tvm.placeholder(shape_c, name="beta", dtype=inp_dtype)
+    gamma_t = tvm.placeholder(shape_c, name="gamma", dtype=inp_dtype)
+    batch_std_t = tvm.placeholder(shape_c, name="batch_std", dtype=inp_dtype)
+    batch_mean_t = tvm.placeholder(shape_c, name="batch_mean", dtype=inp_dtype)
+    running_std_t = tvm.placeholder(shape_c, name="running_std", dtype=inp_dtype)
+
+    res = batchnorm_fold2_compute(x_t, beta_t, gamma_t, batch_std_t, batch_mean_t,
+                                  running_std_t, kernel_name)
+
+    with tvm.target.cce():
+        sch = generic.auto_schedule(res)
+
+    config = {"print_ir": False,
+              "name": kernel_name,
+              "tensor_list": [x_t, beta_t, gamma_t, batch_std_t, batch_mean_t, running_std_t, res]}
+
+    te.lang.cce.cce_build_code(sch, config)

mindspore/ops/_op_impl/_custom_op/batchnorm_fold2_grad.py

@@ -0,0 +1,126 @@
+# Copyright 2020 Huawei Technologies Co., Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ============================================================================
+
+"""_BatchNormFold2Grad op"""
+import te.lang.cce
+from te import tvm
+from te.platform.fusion_manager import fusion_manager
+from topi import generic
+from topi.cce import util
+from mindspore.ops.op_info_register import op_info_register, TBERegOp, DataType
+
+SHAPE_SIZE_LIMIT = 2147483648
+
+batchnorm_fold2_grad_op_info = TBERegOp("BatchNormFold2GradD") \
+    .fusion_type("OPAQUE") \
+    .async_flag(False) \
+    .binfile_name("batchnorm_fold2_grad.so") \
+    .compute_cost(10) \
+    .kernel_name("batchnorm_fold2_grad") \
+    .partial_flag(True) \
+    .op_pattern("formatAgnostic") \
+    .input(0, "dout", None, "required", None) \
+    .input(1, "dout_reduce", None, "required", None) \
+    .input(2, "dout_x_reduce", None, "required", None) \
+    .input(3, "gamma", None, "required", None) \
+    .input(4, "batch_std", None, "required", None) \
+    .input(5, "batch_mean", None, "required", None) \
+    .input(6, "running_std", None, "required", None) \
+    .output(0, "d_batch_std", True, "required", "all") \
+    .output(1, "d_batch_mean", True, "required", "all") \
+    .output(2, "d_gamma", True, "required", "all") \
+    .output(3, "dx", True, "required", "all") \
+    .dtype_format(DataType.F32_5HD, DataType.F32_5HD, DataType.F32_5HD, DataType.F32_5HD, DataType.F32_5HD,
+                  DataType.F32_5HD, DataType.F32_5HD, DataType.F32_5HD, DataType.F32_5HD, DataType.F32_5HD,
+                  DataType.F32_5HD) \
+    .get_op_info()
+
+
+@op_info_register(batchnorm_fold2_grad_op_info)
+def _batchnorm_fold2_grad_tbe():
+    """_BatchNormFold2Grad TBE register"""
+    return
+
+
+@fusion_manager.register("batchnorm_fold2_grad")
+def batchnorm_fold2_grad_compute(dout, dout_reduce, dout_x_reduce, gamma, batch_std, batch_mean, running_std,
+                                 kernel_name="batchnorm_fold2_grad"):
+    """_BatchNormFold2Grad"""
+    shape_x = te.lang.cce.util.shape_to_list(dout.shape)
+
+    d_batch_std_1 = te.lang.cce.vmul(dout_reduce, batch_mean)
+    d_batch_std_1 = te.lang.cce.vmul(d_batch_std_1, gamma)
+    d_batch_std_2 = te.lang.cce.vmul(dout_x_reduce, running_std)
+    d_batch_std = te.lang.cce.vsub(d_batch_std_1, d_batch_std_2)
+    d_batch_std = te.lang.cce.vdiv(d_batch_std, batch_std)
+    d_batch_std = te.lang.cce.vdiv(d_batch_std, batch_std)
+
+    d_batch_mean = te.lang.cce.vmul(dout_reduce, gamma)
+    d_batch_mean = te.lang.cce.vdiv(d_batch_mean, batch_std)
+    d_batch_mean = te.lang.cce.vmuls(d_batch_mean, -1.)
+
+    d_gamma = te.lang.cce.vmul(dout_reduce, batch_mean)
+    d_gamma = te.lang.cce.vdiv(d_gamma, batch_std)
+    d_gamma = te.lang.cce.vmuls(d_gamma, -1.)
+
+    dx = te.lang.cce.vdiv(running_std, batch_std)
+    dx = te.lang.cce.broadcast(dx, shape_x)
+    dx = te.lang.cce.vmul(dx, dout)
+    return [d_batch_std, d_batch_mean, d_gamma, dx]
+
+
+@util.check_input_type(dict, dict, dict, dict, dict, dict, dict, dict, dict, dict, dict, str)
+def batchnorm_fold2_grad(dout, dout_reduce, dout_x_reduce, gamma, batch_std, batch_mean, running_std, d_batch_std,
+                         d_batch_mean, d_gamma, dx, kernel_name="batchnorm_fold2_grad"):
+    """_BatchNormFold2Grad op """
+    shape = dout.get("shape")
+    util.check_kernel_name(kernel_name)
+    util.check_shape_rule(shape)
+    util.check_shape_size(shape, SHAPE_SIZE_LIMIT)
+    check_list = ["float16", "float32"]
+    inp_dtype = dout.get("dtype").lower()
+    if not inp_dtype in check_list:
+        raise RuntimeError("Dtype of input only support float16, float32")
+    data_format = dout.get("format")
+    ori_format = dout.get("ori_format")
+    if data_format.upper() not in ("NC1HWC0", "NCHW"):
+        raise RuntimeError("Un supported data format {}".format(data_format))
+    if data_format.upper() == "NCHW" and ori_format != "NCHW":
+        raise RuntimeError("data_format(NCHW) must same as ori_format")
+    shape_c = gamma.get("shape")
+    if gamma.get("format").upper() == "NCHW":
+        shape_c = 1, gamma.get("shape")[0], 1, 1
+
+    dout_t = tvm.placeholder(shape, name="dout", dtype=inp_dtype)
+    dout_reduce_t = tvm.placeholder(shape_c, name="dout_reduce", dtype=inp_dtype)
+    dout_x_reduce_t = tvm.placeholder(shape_c, name="dout_x_reduce", dtype=inp_dtype)
+    gamma_t = tvm.placeholder(shape_c, name="gamma", dtype=inp_dtype)
+    batch_std_t = tvm.placeholder(shape_c, name="batch_std", dtype=inp_dtype)
+    batch_mean_t = tvm.placeholder(shape_c, name="batch_mean", dtype=inp_dtype)
+    running_std_t = tvm.placeholder(shape_c, name="running_std", dtype=inp_dtype)
+
+    res_list = batchnorm_fold2_grad_compute(dout_t, dout_reduce_t, dout_x_reduce_t, gamma_t, batch_std_t, batch_mean_t,
+                                            running_std_t, kernel_name)
+
+    with tvm.target.cce():
+        sch = generic.auto_schedule(res_list)
+
+    tensor_list = [dout_t, dout_reduce_t, dout_x_reduce_t, gamma_t, batch_std_t, batch_mean_t, running_std_t] + list(
+        res_list)
+    config = {"print_ir": False,
+              "name": kernel_name,
+              "tensor_list": tensor_list}
+
+    te.lang.cce.cce_build_code(sch, config)

mindspore/ops/_op_impl/_custom_op/batchnorm_fold2_grad_reduce.py

@@ -0,0 +1,107 @@
+# Copyright 2020 Huawei Technologies Co., Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ============================================================================
+
+"""_BatchNormFold2GradReduce op"""
+import te.lang.cce
+from te import tvm
+from te.platform.fusion_manager import fusion_manager
+from te.platform.cce_build import build_config
+from topi import generic
+from topi.cce import util
+from mindspore.ops.op_info_register import op_info_register, TBERegOp, DataType
+
+SHAPE_SIZE_LIMIT = 2147483648
+
+batchnorm_fold2_grad_reduce_op_info = TBERegOp("BatchNormFold2GradReduce") \
+    .fusion_type("ELEMWISE") \
+    .async_flag(False) \
+    .binfile_name("batchnorm_fold2_grad_reduce.so") \
+    .compute_cost(10) \
+    .kernel_name("batchnorm_fold2_grad_reduce") \
+    .partial_flag(True) \
+    .op_pattern("formatAgnostic") \
+    .input(0, "dout", None, "required", None) \
+    .input(1, "x", None, "required", None) \
+    .output(0, "dout_reduce", True, "required", "all") \
+    .output(1, "dout_x_reduce", True, "required", "all") \
+    .dtype_format(DataType.F32_5HD, DataType.F32_5HD, DataType.F32_5HD, DataType.F32_5HD) \
+    .get_op_info()
+
+
+@op_info_register(batchnorm_fold2_grad_reduce_op_info)
+def _batchnorm_fold2_grad_reduce_tbe():
+    """_BatchNormFold2GradReduce TBE register"""
+    return
+
+
+@fusion_manager.register("batchnorm_fold2_grad_reduce")
+def batchnorm_fold2_grad_reduce_compute(dout, x, dout_args, kernel_name="batchnorm_fold2_grad_reduce"):
+    """_BatchNormFold2GradReduce compute"""
+    dtype = dout_args.get("dtype")
+    dout_format = dout_args.get("format")
+    ori_format = dout_args.get("ori_format")
+    shape = dout_args.get("shape")
+
+    if dtype == "float16":
+        dout = te.lang.cce.cast_to(dout, "float32")
+        x = te.lang.cce.cast_to(x, "float32")
+
+    dout_x = te.lang.cce.vmul(dout, x)
+    if dout_format == "NC1HWC0":
+        axis = [0, 2, 3]
+        dout_reduce, dout_x_reduce = te.lang.cce.tuple_sum([dout, dout_x], axis, True)
+    else:
+        axis = list(range(len(shape)))
+        if ori_format == "NCHW":
+            axis.pop(1)
+        for _, i in enumerate(range(len(shape))):
+            if shape[i] == 1 and i in axis:
+                axis.remove(i)
+        dout_reduce = te.lang.cce.sum(dout, axis, False)
+        dout_x_reduce = te.lang.cce.sum(dout_x, axis, False)
+    return [dout_reduce, dout_x_reduce]
+
+
+@util.check_input_type(dict, dict, dict, dict, str)
+def batchnorm_fold2_grad_reduce(dout, x, dout_reduce, dout_x_reduce, kernel_name="batchnorm_fold2_grad_reduce"):
+    """_BatchNormFold2GradReduce op"""
+    shape = x.get("shape")
+    x_format = x.get("format")
+    util.check_kernel_name(kernel_name)
+    util.check_shape_rule(shape)
+    util.check_shape_size(shape, SHAPE_SIZE_LIMIT)
+    check_list = ["float16", "float32"]
+    inp_dtype = x.get("dtype").lower()
+    if not inp_dtype in check_list:
+        raise RuntimeError("Dtype of input only support float16, float32")
+    dout_t = tvm.placeholder(shape, name="dout", dtype=inp_dtype)
+    x_t = tvm.placeholder(shape, name="x", dtype=inp_dtype)
+
+    res_list = batchnorm_fold2_grad_reduce_compute(dout_t, x_t, dout, kernel_name)
+
+    if x_format == "NC1HWC0":
+        with tvm.target.cce():
+            sch = generic.auto_schedule(res_list)
+        tensor_list = [dout_t, x_t] + list(res_list)
+        config = {"print_ir": False,
+                  "name": kernel_name,
+                  "tensor_list": tensor_list}
+
+        te.lang.cce.cce_build_code(sch, config)
+        return
+    from impl.bn_training_reduce import bn_training_reduce_schedule_nd
+    sch, tensor_list = bn_training_reduce_schedule_nd(res_list)
+    with build_config:
+        tvm.build(sch, tensor_list, "cce", name=kernel_name)

mindspore/ops/_op_impl/_custom_op/batchnorm_fold_grad.py

@@ -0,0 +1,124 @@
+# Copyright 2020 Huawei Technologies Co., Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ============================================================================
+
+"""_BatchNormFoldGrad op"""
+
+from mindspore.ops.op_info_register import op_info_register, TBERegOp, DataType
+import te.lang.cce
+from te import tvm
+from topi import generic
+from topi.cce import util
+
+batch_norm_op_info = TBERegOp("BatchNormFoldGradD") \
+    .fusion_type("OPAQUE") \
+    .async_flag(False) \
+    .binfile_name("batchnorm_fold_grad.so") \
+    .compute_cost(10) \
+    .kernel_name("batchnorm_fold_grad") \
+    .partial_flag(True) \
+    .attr("epsilon", "optional", "float", "all") \
+    .attr("is_training", "optional", "bool", "all") \
+    .attr("freeze_bn", "optional", "int", "all") \
+    .input(0, "d_batch_mean", False, "required", "all") \
+    .input(1, "d_batch_std", False, "required", "all") \
+    .input(2, "x", False, "required", "all") \
+    .input(3, "batch_mean", False, "required", "all") \
+    .input(4, "batch_std", False, "required", "all") \
+    .output(0, "dx", False, "required", "all") \
+    .dtype_format(DataType.F32_Default, DataType.F32_Default, DataType.F16_Default, DataType.F32_Default,
+                  DataType.F32_Default, DataType.F16_Default) \
+    .dtype_format(DataType.F32_5HD, DataType.F32_5HD, DataType.F16_5HD, DataType.F32_5HD,
+                  DataType.F32_5HD, DataType.F16_5HD) \
+    .dtype_format(DataType.F32_Default, DataType.F32_Default, DataType.F32_Default, DataType.F32_Default,
+                  DataType.F32_Default, DataType.F32_Default) \
+    .dtype_format(DataType.F32_5HD, DataType.F32_5HD, DataType.F32_5HD, DataType.F32_5HD,
+                  DataType.F32_5HD, DataType.F32_5HD) \
+    .get_op_info()
+
+
+@op_info_register(batch_norm_op_info)
+def _batchnorm_fold_grad_tbe():
+    """_BatchNormFoldGrad TBE register"""
+    return
+
+
+def _batchnorm_fold_grad_compute(d_batch_mean, d_batch_std, data_x, batch_mean, batch_std):
+    """_batchnorm_fold_grad_compute """
+    shape_x = te.lang.cce.util.shape_to_list(data_x.shape)
+    normal_size = shape_x[0] * shape_x[2] * shape_x[3]
+
+    d_batch_mean_broad = te.lang.cce.broadcast(d_batch_mean, shape_x)
+    d_batch_std_broad = te.lang.cce.broadcast(d_batch_std, shape_x)
+    batch_mean_broad = te.lang.cce.broadcast(batch_mean, shape_x)
+    batch_std_broad = te.lang.cce.broadcast(batch_std, shape_x)
+
+    dx = te.lang.cce.vsub(data_x, batch_mean_broad)
+    dx = te.lang.cce.vmul(dx, d_batch_std_broad)
+    dx = te.lang.cce.vdiv(dx, batch_std_broad)
+    dx = te.lang.cce.vadd(dx, d_batch_mean_broad)
+    dx = te.lang.cce.vmuls(dx, tvm.const(1. / normal_size, dtype=dx.dtype))
+    return [dx]
+
+
+@util.check_input_type(dict, dict, dict, dict, dict, dict,
+                       float, bool, int, str)
+def batchnorm_fold_grad(d_batch_mean, d_batch_std, x, batch_mean, batch_std, dx,
+                        epsilon=1e-5, is_training=True, freeze_bn=0, kernel_name="batchnorm_fold_grad"):
+    """batchnorm_fold_grad op """
+    util.check_kernel_name(kernel_name)
+    for iv in (d_batch_mean, d_batch_std, x, batch_mean, batch_std):
+        util.check_shape_rule(iv.get("shape"))
+        util.check_tensor_shape_size(iv.get("shape"))
+    check_tuple = ("float16", "float32")
+    for iv in (d_batch_mean, d_batch_std, x, batch_mean, batch_std):
+        util.check_dtype_rule(iv.get("dtype").lower(), check_tuple)
+
+    shape_x = x.get("shape")
+    dtype_x = x.get("dtype")
+    format_data = x.get("format").upper()
+    if format_data not in ("NCHW", "NC1HWC0"):
+        raise RuntimeError("Format of input only support 4D and 5HD")
+
+    shape_mean = d_batch_mean.get("shape")
+    dtype_mean = d_batch_mean.get("dtype").lower()
+    if format_data == "NC1HWC0":
+        if len(shape_x) != 5:
+            raise RuntimeError("batchnorm_fold only support shape 5D"
+                               "when input format is NC1HWC0")
+        shape_mean = (1, shape_x[1], 1, 1, shape_x[4])
+    elif format_data == "NCHW":
+        if len(shape_x) < 2 or len(shape_x) > 4:
+            raise RuntimeError("batchnorm_fold only support shape 2D to 4D")
+        if shape_x[1] != shape_mean[0]:
+            raise RuntimeError("data_format is NCHW, shape_bias must"
+                               "be equal to the second axis of shape_x")
+        shape_mean = (1, shape_x[1],)
+        for _ in range(2, len(shape_x)):
+            shape_mean = shape_mean + (1,)
+
+    d_batch_mean = tvm.placeholder(shape_mean, name="d_batch_mean", dtype=dtype_mean)
+    d_batch_std = tvm.placeholder(shape_mean, name="d_batch_std", dtype=dtype_mean)
+    data_x = tvm.placeholder(shape_x, name="data_x", dtype=dtype_x.lower())
+    batch_mean = tvm.placeholder(shape_mean, name="batch_mean", dtype=dtype_mean)
+    batch_std = tvm.placeholder(shape_mean, name="batch_std", dtype=dtype_mean)
+
+    res = _batchnorm_fold_grad_compute(d_batch_mean, d_batch_std, data_x, batch_mean, batch_std)
+    with tvm.target.cce():
+        sch = generic.auto_schedule(res)
+
+    tensor_list = [d_batch_mean, d_batch_std, data_x, batch_mean, batch_std] + res
+    config = {"name": kernel_name,
+              "tensor_list": tensor_list}
+    te.lang.cce.cce_build_code(sch, config)

mindspore/ops/_op_impl/_custom_op/correction_mul.py

@@ -0,0 +1,92 @@
+# Copyright 2020 Huawei Technologies Co., Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ============================================================================
+
+"""CorrectionMul op"""
+import te.lang.cce
+from te import tvm
+from te.platform.fusion_manager import fusion_manager
+from topi import generic
+from topi.cce import util
+from mindspore.ops.op_info_register import op_info_register, TBERegOp, DataType
+
+SHAPE_SIZE_LIMIT = 2147483648
+
+correction_mul_op_info = TBERegOp("CorrectionMul") \
+    .fusion_type("ELEMWISE") \
+    .async_flag(False) \
+    .binfile_name("correction_mul.so") \
+    .compute_cost(10) \
+    .kernel_name("correction_mul") \
+    .partial_flag(True) \
+    .op_pattern("formatAgnostic") \
+    .attr("channel_axis", "optional", "int", "all") \
+    .input(0, "x", None, "required", None) \
+    .input(1, "batch_std", None, "required", None) \
+    .input(2, "running_std", None, "required", None) \
+    .output(0, "y", True, "required", "all") \
+    .dtype_format(DataType.F16_Default, DataType.F16_Default, DataType.F16_Default, DataType.F16_Default) \
+    .dtype_format(DataType.F16_5HD, DataType.F16_5HD, DataType.F16_5HD, DataType.F16_5HD) \
+    .dtype_format(DataType.F32_Default, DataType.F32_Default, DataType.F32_Default, DataType.F32_Default) \
+    .dtype_format(DataType.F32_5HD, DataType.F32_5HD, DataType.F32_5HD, DataType.F32_5HD) \
+    .get_op_info()
+
+
+@op_info_register(correction_mul_op_info)
+def _correction_mul_tbe():
+    """CorrectionMul TBE register"""
+    return
+
+
+@fusion_manager.register("correction_mul")
+def correction_mul_compute(x, batch_std, running_std, kernel_name="correction_mul"):
+    """CorrectionMul compute"""
+    shape_x = te.lang.cce.util.shape_to_list(x.shape)
+    factor = te.lang.cce.vdiv(batch_std, running_std)
+    factor_b = te.lang.cce.broadcast(factor, shape_x)
+    res = te.lang.cce.vmul(x, factor_b)
+    return res
+
+
+@util.check_input_type(dict, dict, dict, dict, int, str)
+def correction_mul(x, batch_std, running_std, y, channel, kernel_name="correction_mul"):
+    """CorrectionMul op"""
+    shape = x.get("shape")
+    data_format = x.get("format")
+    util.check_kernel_name(kernel_name)
+    util.check_shape_rule(shape)
+    util.check_shape_size(shape, SHAPE_SIZE_LIMIT)
+    check_list = ["float16", "float32"]
+    inp_dtype = x.get("dtype").lower()
+    if not inp_dtype in check_list:
+        raise RuntimeError("Dtype of input only support float16, float32")
+
+    # shape = util.shape_refine(shape)
+    x_t = tvm.placeholder(shape, name="x", dtype=inp_dtype)
+    shape_c = [1] * len(shape)
+    shape_c[channel] = batch_std.get("ori_shape")[0]
+    if data_format == "NC1HWC0" and channel == 1:
+        shape_c = batch_std.get("shape")
+    batch_std_t = tvm.placeholder(shape_c, name="batch_std", dtype=inp_dtype)
+    running_std_t = tvm.placeholder(shape_c, name="running_std", dtype=inp_dtype)
+    res = correction_mul_compute(x_t, batch_std_t, running_std_t, kernel_name)
+
+    with tvm.target.cce():
+        sch = generic.auto_schedule(res)
+
+    config = {"print_ir": False,
+              "name": kernel_name,
+              "tensor_list": [x_t, batch_std_t, running_std_t, res]}
+
+    te.lang.cce.cce_build_code(sch, config)

mindspore/ops/_op_impl/_custom_op/correction_mul_grad.py

@@ -0,0 +1,134 @@
+# Copyright 2020 Huawei Technologies Co., Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ============================================================================
+
+"""CorrectionMul op"""
+import te.lang.cce
+from te import tvm
+from te.platform.fusion_manager import fusion_manager
+from topi import generic
+from topi.cce import util
+from mindspore.ops.op_info_register import op_info_register, TBERegOp, DataType
+
+SHAPE_SIZE_LIMIT = 2147483648
+
+correction_mul_grad_op_info = TBERegOp("CorrectionMulGrad") \
+    .fusion_type("OPAQUE") \
+    .async_flag(False) \
+    .binfile_name("correction_mul_grad.so") \
+    .compute_cost(10) \
+    .kernel_name("correction_mul_grad") \
+    .partial_flag(True) \
+    .op_pattern("formatAgnostic") \
+    .attr("channel_axis", "optional", "int", "all") \
+    .input(0, "dout", None, "required", None) \
+    .input(1, "x", None, "required", None) \
+    .input(2, "batch_std", None, "required", None) \
+    .input(3, "running_std", None, "required", None) \
+    .output(0, "dx", True, "required", "all") \
+    .output(1, "d_batch_std", True, "required", "all") \
+    .dtype_format(DataType.F16_Default, DataType.F16_Default, DataType.F16_Default, DataType.F16_Default,
+                  DataType.F16_Default, DataType.F16_Default) \
+    .dtype_format(DataType.F16_5HD, DataType.F16_5HD, DataType.F16_5HD, DataType.F16_5HD,
+                  DataType.F16_5HD, DataType.F16_5HD) \
+    .dtype_format(DataType.F32_Default, DataType.F32_Default, DataType.F32_Default, DataType.F32_Default,
+                  DataType.F32_Default, DataType.F32_Default) \
+    .dtype_format(DataType.F32_5HD, DataType.F32_5HD, DataType.F32_5HD, DataType.F32_5HD,
+                  DataType.F32_5HD, DataType.F32_5HD) \
+    .get_op_info()
+
+
+@op_info_register(correction_mul_grad_op_info)
+def _correction_mul_grad_tbe():
+    """CorrectionMulGrad TBE register"""
+    return
+
+
+@fusion_manager.register("correction_mul_grad")
+def correction_mul_grad_compute(dout, x, batch_std, running_std, channel, data_format, kernel_name="correction_mul"):
+    """CorrectionMulGrad compute"""
+    shape_x = te.lang.cce.util.shape_to_list(x.shape)
+    factor = te.lang.cce.vdiv(batch_std, running_std)
+    factor_b = te.lang.cce.broadcast(factor, shape_x)
+    dx = te.lang.cce.vmul(dout, factor_b)
+    mul_data = te.lang.cce.vmul(dout, x)
+    if channel == 0:
+        if data_format == "NCHW":
+            axis = [1, 2, 3]
+        else:
+            axis = [1, 2, 3, 4]
+    else:
+        axis = [2, 3]
+    red_data = te.lang.cce.sum(mul_data, axis, keepdims=True)
+    d_batch_std = te.lang.cce.vdiv(red_data, running_std)
+    return [dx, d_batch_std]
+
+
+@util.check_input_type(dict, dict, dict, dict, dict, dict, int, str)
+def correction_mul_grad(dout, x, batch_std, running_std, dx, d_batch_std, channel, kernel_name="correction_mul_grad"):
+    """CorrectionMulGrad op"""
+    shape_dout = dout.get("shape")
+    shape_x = dout.get("shape")
+
+    dtype_dout = dout.get("dtype")
+    dtype_x = x.get("dtype")
+    dtype_batch_std = batch_std.get("dtype")
+    dtype_running_std = running_std.get("dtype")
+
+    inp_dtype_dout = dtype_dout.lower()
+    inp_dtype_x = dtype_x.lower()
+    inp_dtype_batch_std = dtype_batch_std.lower()
+    inp_dtype_running_std = dtype_running_std.lower()
+
+    util.check_dtype_rule(inp_dtype_dout, ("float16", "float32"))
+    util.check_dtype_rule(inp_dtype_x, ("float16", "float32"))
+    util.check_dtype_rule(inp_dtype_batch_std, ("float32",))
+    util.check_dtype_rule(inp_dtype_running_std, ("float32",))
+    util.compare_tensor_dict_key(dout, x, "dtype")
+    util.compare_tensor_dict_key(dout, x, "shape")
+    util.compare_tensor_dict_key(dx, x, "shape")
+    util.compare_tensor_dict_key(batch_std, running_std, "shape")
+    util.compare_tensor_dict_key(batch_std, d_batch_std, "shape")
+
+    util.check_kernel_name(kernel_name)
+    util.check_shape_rule(shape_x)
+    util.check_shape_size(shape_x, SHAPE_SIZE_LIMIT)
+
+    data_format = dout.get("format")
+    ori_format = dout.get("format")
+    if data_format.upper() not in ("NC1HWC0", "NCHW"):
+        raise RuntimeError("Un supported data format {}".format(data_format))
+    if data_format.upper() == "NCHW" and ori_format != "NCHW":
+        raise RuntimeError("data_format(NCHW) must same as ori_format")
+
+    shape_c = [1] * len(shape_x)
+    shape_c[channel] = batch_std.get("ori_shape")[0]
+    if data_format == "NC1HWC0" and channel == 1:
+        shape_c = batch_std.get("shape")
+
+    dout_t = tvm.placeholder(shape_dout, name="dout", dtype=inp_dtype_dout)
+    x_t = tvm.placeholder(shape_x, name="x", dtype=inp_dtype_x)
+    batch_std_t = tvm.placeholder(shape_c, name="batch_std", dtype=inp_dtype_batch_std)
+    running_std_t = tvm.placeholder(shape_c, name="running_std", dtype=inp_dtype_running_std)
+    res_list = correction_mul_grad_compute(dout_t, x_t, batch_std_t, running_std_t, channel, data_format, kernel_name)
+
+    with tvm.target.cce():
+        sch = generic.auto_schedule(res_list)
+
+    tensor_list = [dout_t, x_t, batch_std_t, running_std_t] + list(res_list)
+    config = {"print_ir": False,
+              "name": kernel_name,
+              "tensor_list": tensor_list}
+
+    te.lang.cce.cce_build_code(sch, config)

mindspore/ops/_op_impl/_custom_op/fake_quant_with_min_max.py

@@ -0,0 +1,146 @@
+# Copyright 2020 Huawei Technologies Co., Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ============================================================================
+
+"""FakeQuantWithMinMax op"""
+
+from functools import reduce as functools_reduce
+import te.lang.cce
+from te import tvm
+from te.platform.fusion_manager import fusion_manager
+from topi import generic
+from topi.cce import util
+from mindspore.ops.op_info_register import op_info_register, TBERegOp, DataType
+
+fake_quant_op_info = TBERegOp("FakeQuantWithMinMax") \
+    .fusion_type("ELEMWISE") \
+    .async_flag(False) \
+    .binfile_name("fake_quant_with_min_max_vars_ema.so") \
+    .compute_cost(10) \
+    .kernel_name("fake_quant_with_min_max_vars_ema") \
+    .partial_flag(True) \
+    .attr("ema", "optional", "bool", "all") \
+    .attr("ema_decay", "optional", "float", "all") \
+    .attr("symmetric", "optional", "bool", "all") \
+    .attr("narrow_range", "optional", "bool", "all") \
+    .attr("training", "optional", "bool", "all") \
+    .attr("num_bits", "optional", "int", "all") \
+    .attr("quant_delay", "optional", "int", "all") \
+    .input(0, "x", None, "required", None) \
+    .input(1, "min", None, "required", None) \
+    .input(2, "max", None, "required", None) \
+    .output(0, "y", True, "required", "all") \
+    .dtype_format(DataType.F16_Default, DataType.F16_Default, DataType.F16_Default, DataType.F16_Default) \
+    .dtype_format(DataType.F16_5HD, DataType.F16_5HD, DataType.F16_5HD, DataType.F16_5HD) \
+    .dtype_format(DataType.F32_Default, DataType.F32_Default, DataType.F32_Default, DataType.F32_Default) \
+    .dtype_format(DataType.F32_5HD, DataType.F32_5HD, DataType.F32_5HD, DataType.F32_5HD) \
+    .get_op_info()
+
+
+@op_info_register(fake_quant_op_info)
+def _fake_quant_tbe():
+    """FakeQuantWithMinMax TBE register"""
+    return
+
+
+@fusion_manager.register("fake_quant_with_min_max_vars_ema")
+def fake_quant_with_min_max_vars_ema_compute(x, min_val, max_val, y, quant_min, quant_max,
+                                             kernel_name="correction_mul"):
+    """FakeQuantWithMinMax"""
+    shape = te.lang.cce.util.shape_to_list(x.shape)
+    shape_min = te.lang.cce.util.shape_to_list(min_val.shape)
+    quant_min = te.lang.cce.broadcast(quant_min, shape_min, x.dtype)
+    quant_max = te.lang.cce.broadcast(quant_max, shape_min, x.dtype)
+    min_val = te.lang.cce.broadcast(min_val, shape_min, x.dtype)
+    max_val = te.lang.cce.broadcast(max_val, shape_min, x.dtype)
+
+    # CalNudge(NudgeMinMax)
+    scale = te.lang.cce.vdiv(te.lang.cce.vsub(max_val, min_val), te.lang.cce.vsub(quant_max, quant_min))
+    zp_from_min = te.lang.cce.vsub(quant_min, te.lang.cce.vdiv(min_val, scale))
+    # Nudge zero point
+    nudge_zp = te.lang.cce.round(te.lang.cce.vmin(quant_max, te.lang.cce.vmax(quant_min, zp_from_min)))
+    nudge_min = te.lang.cce.vmul(te.lang.cce.vsub(quant_min, nudge_zp), scale)
+    nudge_max = te.lang.cce.vmul(te.lang.cce.vsub(quant_max, nudge_zp), scale)
+
+    # boradcast to shape
+    nudge_min = te.lang.cce.broadcast(nudge_min, shape, x.dtype)
+    nudge_max = te.lang.cce.broadcast(nudge_max, shape, x.dtype)
+    scale = te.lang.cce.broadcast(scale, shape, x.dtype)
+
+    # FakeQuant
+    input_x = te.lang.cce.vmin(nudge_max, te.lang.cce.vmax(nudge_min, x))
+    nudge_input = te.lang.cce.floor(te.lang.cce.vadds(te.lang.cce.vdiv(te.lang.cce.vsub(input_x, nudge_min), scale),
+                                                      0.5))
+    res = te.lang.cce.vadd(te.lang.cce.vmul(nudge_input, scale), nudge_min)
+
+    return res
+
+
+@util.check_input_type(dict, dict, dict, dict, bool, float, bool, bool, bool, int, int, str)
+def fake_quant_with_min_max_vars_ema(x, min_val, max_val, y,
+                                     ema, ema_decay, symmetric, narrow_range, training, num_bits, quant_delay,
+                                     kernel_name="fake_quant"):
+    """FakeQuantWithMinMax"""
+    input_shape = x.get("shape")
+    input_dtype = x.get("dtype")
+    min_shape = min_val.get("ori_shape")
+    min_dtype = min_val.get("dtype")
+    max_shape = max_val.get("ori_shape")
+    max_dtype = max_val.get("dtype")
+
+    min_shape = util.scalar2tensor_one(min_shape)
+    max_shape = util.scalar2tensor_one(max_shape)
+    util.check_kernel_name(kernel_name)
+    util.check_shape_rule(input_shape)
+    util.check_shape_rule(min_shape, 1, 1, 1)
+    util.check_shape_rule(max_shape, 1, 1, 1)
+    util.check_tensor_shape_size(input_shape)
+    util.check_tensor_shape_size(min_shape)
+    util.check_tensor_shape_size(max_shape)
+
+    check_list = ["float32", "float16"]
+    x_dtype = input_dtype.lower()
+    min_dtype = min_dtype.lower()
+    max_dtype = max_dtype.lower()
+    util.check_dtype_rule(x_dtype, check_list)
+    util.check_dtype_rule(min_dtype, check_list)
+    util.check_dtype_rule(max_dtype, check_list)
+
+    input_shape = (functools_reduce(lambda x, y: x * y, input_shape[:]),)
+    shape_min, _, _ = util.produce_shapes(min_shape, input_shape)
+
+    if symmetric:
+        quant_min = 0 - 2 ** (num_bits - 1)
+        quant_max = 2 ** (num_bits - 1) - 1
+    else:
+        quant_min = 0
+        quant_max = 2 ** num_bits - 1
+    if narrow_range:
+        quant_min = quant_min + 1
+
+    input_data = tvm.placeholder(input_shape, name="x", dtype=x_dtype)
+    min_data = tvm.placeholder(shape_min, name="min_data", dtype=min_dtype)
+    max_data = tvm.placeholder(shape_min, name="max_data", dtype=max_dtype)
+    res = fake_quant_with_min_max_vars_ema_compute(input_data, min_data, max_data, y,
+                                                   quant_min, quant_max, kernel_name)
+
+    with tvm.target.cce():
+        sch = generic.auto_schedule(res)
+
+    tensor_list = [input_data, min_data, max_data, res]
+    config = {"print_ir": False,
+              "name": kernel_name,
+              "tensor_list": tensor_list}
+
+    te.lang.cce.cce_build_code(sch, config)

mindspore/ops/_op_impl/_custom_op/fake_quant_with_min_max_grad.py

@@ -0,0 +1,156 @@
+# Copyright 2020 Huawei Technologies Co., Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ============================================================================
+
+"""FakeQuantWithMinMaxGrad op"""
+
+from functools import reduce as functools_reduce
+import te.lang.cce
+from te import tvm
+from te.platform.fusion_manager import fusion_manager
+from topi import generic
+from topi.cce import util
+from mindspore.ops.op_info_register import op_info_register, TBERegOp, DataType
+
+SHAPE_SIZE_LIMIT = 2147483648
+D_TYPE = 'float32'
+
+fake_quant_grad_op_info = TBERegOp("FakeQuantWithMinMaxGrad") \
+    .fusion_type("OPAQUE") \
+    .async_flag(False) \
+    .binfile_name("fake_quant_with_min_max_grad.so") \
+    .compute_cost(10) \
+    .kernel_name("fake_quant_with_min_max_grad") \
+    .partial_flag(True) \
+    .attr("num_bits", "optional", "int", "all") \
+    .attr("quant_delay", "optional", "int", "all") \
+    .input(0, "dout", None, "required", None) \
+    .input(1, "x", None, "required", None) \
+    .input(2, "min", None, "required", None) \
+    .input(3, "max", None, "required", None) \
+    .output(0, "dx", True, "required", "all") \
+    .dtype_format(DataType.F16_Default, DataType.F16_Default, DataType.F16_Default, DataType.F16_Default,
+                  DataType.F16_Default) \
+    .dtype_format(DataType.F16_5HD, DataType.F16_5HD, DataType.F16_5HD, DataType.F16_5HD, DataType.F16_5HD) \
+    .dtype_format(DataType.F32_Default, DataType.F32_Default, DataType.F32_Default, DataType.F32_Default,
+                  DataType.F32_Default) \
+    .dtype_format(DataType.F32_5HD, DataType.F32_5HD, DataType.F32_5HD, DataType.F32_5HD, DataType.F32_5HD) \
+    .get_op_info()
+
+
+def _less_compare_float32(data_x, data_y):
+    """_less_compare_float32 compute"""
+    shape_inputs = te.lang.cce.util.shape_to_list(data_x.shape)
+    min_value = tvm.const(2 ** (-126), dtype=D_TYPE)
+    max_value = tvm.const(2 ** 62, dtype=D_TYPE)
+    factor_value = tvm.const(2 ** 2, dtype=D_TYPE)
+    data_zero = te.lang.cce.broadcast(tvm.const(0, dtype=D_TYPE), shape_inputs, D_TYPE)
+    min_value_tensor = te.lang.cce.vadds(data_zero, min_value)
+
+    res_sub = te.lang.cce.vsub(data_y, data_x)
+    res_min = te.lang.cce.vmin(res_sub, min_value_tensor)
+    res_max = te.lang.cce.vmax(res_min, data_zero)
+
+    res_max_mul = te.lang.cce.vmuls(res_max, max_value)
+    res_max_mul_max = te.lang.cce.vmuls(res_max_mul, max_value)
+    res = te.lang.cce.vmuls(res_max_mul_max, factor_value)
+
+    return res
+
+
+@op_info_register(fake_quant_grad_op_info)
+def _fake_quant_grad_tbe():
+    """FakeQuantWithMinMaxGrad TBE register"""
+    return
+
+
+@fusion_manager.register("fake_quant_with_min_max_grad")
+def fake_quant_with_min_max_grad_compute(dout, x, min_val, max_val, quant_min, quant_max,
+                                         kernel_name="fake_quant_with_min_max_grad"):
+    """FakeQuantWithMinMaxGrad"""
+    shape = te.lang.cce.util.shape_to_list(x.shape)
+    shape_min = te.lang.cce.util.shape_to_list(min_val.shape)
+    quant_min = tvm.const(quant_min, x.dtype)
+    quant_max = tvm.const(quant_max, x.dtype)
+    quant_min = te.lang.cce.broadcast(quant_min, shape_min)
+    quant_max = te.lang.cce.broadcast(quant_max, shape_min)
+
+    # CalNudge(NudgeMinMax)
+    scale = te.lang.cce.vdiv(te.lang.cce.vsub(max_val, min_val), te.lang.cce.vsub(quant_max, quant_min))
+    zp_from_min = te.lang.cce.vsub(quant_min, te.lang.cce.vdiv(min_val, scale))
+    # Nudge zero point
+    nudge_zp = te.lang.cce.round(te.lang.cce.vmin(quant_max, te.lang.cce.vmax(quant_min, zp_from_min)))
+    nudge_min = te.lang.cce.vmul(te.lang.cce.vsub(quant_min, nudge_zp), scale)
+    nudge_max = te.lang.cce.vmul(te.lang.cce.vsub(quant_max, nudge_zp), scale)
+    nudge_min = te.lang.cce.broadcast(nudge_min, shape)
+    nudge_max = te.lang.cce.broadcast(nudge_max, shape)
+
+    bool_over_min = _less_compare_float32(nudge_min, x)
+    bool_less_max = _less_compare_float32(x, nudge_max)
+    bool_between = te.lang.cce.vmul(bool_over_min, bool_less_max)
+    res = te.lang.cce.vmul(dout, bool_between)
+
+    return res
+
+
+@util.check_input_type(dict, dict, dict, dict, dict, int, int, str)
+def fake_quant_with_min_max_grad(dout, x, min_val, max_val, dx, num_bits, quant_delay,
+                                 kernel_name="fake_quant_with_min_max_grad"):
+    """FakeQuantWithMinMaxGrad"""
+    input_shape = x.get("shape")
+    input_dtype = x.get("dtype")
+    min_shape = min_val.get("ori_shape")
+    min_dtype = min_val.get("dtype")
+    max_shape = max_val.get("ori_shape")
+    max_dtype = max_val.get("dtype")
+
+    min_shape = util.scalar2tensor_one(min_shape)
+    max_shape = util.scalar2tensor_one(max_shape)
+    util.check_kernel_name(kernel_name)
+    util.check_shape_rule(input_shape)
+    util.check_shape_rule(min_shape, 1, 1, 1)
+    util.check_shape_rule(max_shape, 1, 1, 1)
+    util.check_tensor_shape_size(input_shape)
+    util.check_tensor_shape_size(min_shape)
+    util.check_tensor_shape_size(max_shape)
+
+    check_list = ["float32", 'float16']
+    x_dtype = input_dtype.lower()
+    min_dtype = min_dtype.lower()
+    max_dtype = max_dtype.lower()
+    util.check_dtype_rule(x_dtype, check_list)
+    util.check_dtype_rule(min_dtype, check_list)
+    util.check_dtype_rule(max_dtype, check_list)
+
+    input_shape = (functools_reduce(lambda x, y: x * y, input_shape[:]),)
+    shape_min, _, _ = util.produce_shapes(min_shape, input_shape)
+
+    quant_min = 0
+    quant_max = 2 ** num_bits - 1
+    dout_data = tvm.placeholder(input_shape, name="dout", dtype=x_dtype)
+    input_data = tvm.placeholder(input_shape, name="x", dtype=x_dtype)
+    min_data = tvm.placeholder(shape_min, name="min_data", dtype=min_dtype)
+    max_data = tvm.placeholder(shape_min, name="max_data", dtype=max_dtype)
+    res = fake_quant_with_min_max_grad_compute(dout_data, input_data, min_data, max_data, quant_min,
+                                               quant_max, kernel_name)
+
+    with tvm.target.cce():
+        sch = generic.auto_schedule(res)
+
+    tensor_list = [dout_data, input_data, min_data, max_data, res]
+    config = {"print_ir": False,
+              "name": kernel_name,
+              "tensor_list": tensor_list}
+
+    te.lang.cce.cce_build_code(sch, config)

mindspore/ops/_op_impl/_custom_op/fake_quant_with_min_max_update.py

@@ -0,0 +1,137 @@
+# Copyright 2020 Huawei Technologies Co., Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ============================================================================
+
+"""FakeQuantWithMinMaxUpdate op"""
+from functools import reduce as functools_reduce
+import te.lang.cce
+from te import tvm
+from te.platform.fusion_manager import fusion_manager
+from topi import generic
+from topi.cce import util
+from mindspore.ops.op_info_register import op_info_register, TBERegOp, DataType
+
+
+fake_quant_update5d_op_info = TBERegOp("FakeQuantWithMinMaxUpdate") \
+    .fusion_type("OPAQUE") \
+    .async_flag(False) \
+    .binfile_name("fake_quant_with_min_max_update5d.so") \
+    .compute_cost(10) \
+    .kernel_name("fake_quant_with_min_max_update") \
+    .partial_flag(True) \
+    .attr("ema", "optional", "bool", "all") \
+    .attr("ema_decay", "optional", "float", "all") \
+    .attr("symmetric", "optional", "bool", "all") \
+    .attr("narrow_range", "optional", "bool", "all") \
+    .attr("training", "optional", "bool", "all") \
+    .attr("num_bits", "optional", "int", "all") \
+    .attr("quant_delay", "optional", "int", "all") \
+    .input(0, "x", None, "required", None) \
+    .input(1, "min", None, "required", None) \
+    .input(2, "max", None, "required", None) \
+    .output(0, "min_up", True, "required", "all") \
+    .output(1, "max_up", True, "required", "all") \
+    .dtype_format(DataType.F32_5HD, DataType.F32_5HD, DataType.F32_5HD, DataType.F32_5HD,
+                  DataType.F32_5HD) \
+    .get_op_info()
+
+
+@op_info_register(fake_quant_update5d_op_info)
+def _fake_quant_update5d_tbe():
+    """_FakeQuantWithMinMaxUpdate5D TBE register"""
+    return
+
+
+@fusion_manager.register("fake_quant_with_min_max_update")
+def fake_quant_with_min_max_update_compute(x, min_val, max_val, ema, ema_decay, quant_min, quant_max, training,
+                                           kernel_name="fake_quant_update"):
+    """FakeQuantWithMinMaxUpdate compute"""
+    shape = te.lang.cce.util.shape_to_list(x.shape)
+    shape_min = te.lang.cce.util.shape_to_list(min_val.shape)
+    min_val = te.lang.cce.broadcast(min_val, shape_min, x.dtype)
+    max_val = te.lang.cce.broadcast(max_val, shape_min, x.dtype)
+    if not ema:
+        ema_decay = 0.0
+    if training:
+        # CalMinMax
+        axis = tuple(range(len(shape)))
+        x_min = te.lang.cce.reduce_min(x, axis=axis)
+        x_max = te.lang.cce.reduce_max(x, axis=axis)
+        x_min = te.lang.cce.broadcast(x_min, shape_min)
+        x_max = te.lang.cce.broadcast(x_max, shape_min)
+        min_val = te.lang.cce.vadd(te.lang.cce.vmuls(min_val, ema_decay), te.lang.cce.vmuls(x_min, (1 - ema_decay)))
+        max_val = te.lang.cce.vadd(te.lang.cce.vmuls(max_val, ema_decay), te.lang.cce.vmuls(x_max, (1 - ema_decay)))
+        min_val = te.lang.cce.vmins(min_val, 0)
+        max_val = te.lang.cce.vmaxs(max_val, 0)
+
+    return [min_val, max_val]
+
+
+@util.check_input_type(dict, dict, dict, dict, dict, bool, float, bool, bool, bool, int, int, str)
+def fake_quant_with_min_max_update(x, min_val, max_val, min_up, max_up,
+                                   ema, ema_decay, symmetric, narrow_range, training, num_bits, quant_delay,
+                                   kernel_name="fake_quant_update"):
+    """FakeQuantWithMinMax op"""
+    input_shape = x.get("shape")
+    input_dtype = x.get("dtype")
+    min_shape = min_val.get("ori_shape")
+    min_dtype = min_val.get("dtype")
+    max_shape = max_val.get("ori_shape")
+    max_dtype = max_val.get("dtype")
+
+    min_shape = util.scalar2tensor_one(min_shape)
+    max_shape = util.scalar2tensor_one(max_shape)
+    util.check_kernel_name(kernel_name)
+    util.check_shape_rule(input_shape)
+    util.check_shape_rule(min_shape, 1, 1, 1)
+    util.check_shape_rule(max_shape, 1, 1, 1)
+    util.check_tensor_shape_size(input_shape)
+    util.check_tensor_shape_size(min_shape)
+    util.check_tensor_shape_size(max_shape)
+
+    check_list = ["float32", "float16"]
+    x_dtype = input_dtype.lower()
+    min_dtype = min_dtype.lower()
+    max_dtype = max_dtype.lower()
+    util.check_dtype_rule(x_dtype, check_list)
+    util.check_dtype_rule(min_dtype, check_list)
+    util.check_dtype_rule(max_dtype, check_list)
+
+    input_shape = (functools_reduce(lambda x, y: x * y, input_shape[:]),)
+    shape_min, _, _ = util.produce_shapes(min_shape, input_shape)
+
+    if symmetric:
+        quant_min = 0 - 2 ** (num_bits - 1)
+        quant_max = 2 ** (num_bits - 1) - 1
+    else:
+        quant_min = 0
+        quant_max = 2 ** num_bits - 1
+    if narrow_range:
+        quant_min = quant_min + 1
+
+    input_data = tvm.placeholder(input_shape, name="x", dtype=x_dtype)
+    min_data = tvm.placeholder(shape_min, name="min_data", dtype=min_dtype)
+    max_data = tvm.placeholder(shape_min, name="max_data", dtype=max_dtype)
+    res_list = fake_quant_with_min_max_update_compute(input_data, min_data, max_data,
+                                                      ema, ema_decay, quant_min, quant_max, training, kernel_name)
+
+    with tvm.target.cce():
+        sch = generic.auto_schedule(res_list)
+
+    tensor_list = [input_data, min_data, max_data] + list(res_list)
+    config = {"print_ir": False,
+              "name": kernel_name,
+              "tensor_list": tensor_list}
+
+    te.lang.cce.cce_build_code(sch, config)

mindspore/ops/operations/_quant_ops.py

@@ -30,6 +30,10 @@ __all__ = ["FakeQuantWithMinMax",
            "CorrectionMulGrad",
            "BatchNormFold2",
            "BatchNormFold2Grad",
+           "BatchNormFoldD",
+           "BNTrainingReduce",
+           "BatchNormFold2_D",
+           "FakeQuantWithMinMaxUpdate",
            ]
 
 
@@ -166,7 +170,7 @@ class FakeQuantWithMinMaxPerChannel(PrimitiveWithInfer):
         >>> result = fake_quant(input_x, _min, _max)
     """
     support_quant_bit = [4, 8]
-    channel_idx = 0
+    channel_axis = 0
 
     @prim_attr_register
     def __init__(self, num_bits=8, ema=False, ema_decay=0.999, quant_delay=0, symmetric=False, narrow_range=False,
@@ -188,8 +192,8 @@ class FakeQuantWithMinMaxPerChannel(PrimitiveWithInfer):
 
     def infer_shape(self, x_shape, min_shape, max_shape):
         validator.check_integer("x rank", len(x_shape), 1, Rel.GT, self.name)
-        validator.check_integer("min shape[0]", min_shape[0], x_shape[self.channel_idx], Rel.EQ, self.name)
+        validator.check_integer("min shape[0]", min_shape[0], x_shape[self.channel_axis], Rel.EQ, self.name)
-        validator.check_integer("max shape[0]", max_shape[0], x_shape[self.channel_idx], Rel.EQ, self.name)
+        validator.check_integer("max shape[0]", max_shape[0], x_shape[self.channel_axis], Rel.EQ, self.name)
         return x_shape
 
     def infer_dtype(self, x_type, min_type, max_type):
@@ -272,7 +276,7 @@ class BatchNormFold(PrimitiveWithInfer):
         >>> global_step = Tensor(np.arange(6), mindspore.int32)
         >>> batch_mean, batch_std, running_mean, running_std = batch_norm_fold(input_x, mean, variance, global_step)
     """
-    channel = 1
+    channel_axis = 1
 
     @prim_attr_register
     def __init__(self, momentum=0.1, epsilon=1e-5, is_training=True, freeze_bn=0):
@@ -287,7 +291,7 @@ class BatchNormFold(PrimitiveWithInfer):
 
     def infer_shape(self, x_shape, mean_shape, variance_shape, global_step_shape):
         validator.check("mean shape", mean_shape, "gamma_shape", variance_shape, Rel.EQ, self.name)
-        validator.check("mean_shape[0]", mean_shape[0], "input channel", x_shape[self.channel], Rel.EQ, self.name)
+        validator.check("mean_shape[0]", mean_shape[0], "input channel", x_shape[self.channel_axis], Rel.EQ, self.name)
         validator.check_integer("global_step rank", len(global_step_shape), 1, Rel.EQ, self.name)
         return mean_shape, mean_shape, mean_shape, mean_shape
 
@@ -314,7 +318,7 @@ class BatchNormFoldGrad(PrimitiveWithInfer):
         >>> global_step = Tensor([2], mindspore.int32)
         >>> result = batch_norm_fold_grad(d_batch_mean, d_batch_std, input_x, batch_mean, batch_std, global_step)
     """
-    channel = 1
+    channel_axis = 1
 
     @prim_attr_register
     def __init__(self, epsilon=1e-5, is_training=True, freeze_bn=0):
@@ -333,8 +337,8 @@ class BatchNormFoldGrad(PrimitiveWithInfer):
                         "batch_mean shape", batch_mean_shape, Rel.EQ, self.name)
         validator.check("d_batch_mean shape", d_batch_mean_shape,
                         "batch_std shape", batch_std_shape, Rel.EQ, self.name)
-        validator.check("d_batch_mean_shape[0]", d_batch_mean_shape[0], "input channel", x_shape[self.channel], Rel.EQ,
+        validator.check("d_batch_mean_shape[0]", d_batch_mean_shape[0],
-                        self.name)
+                        "input channel", x_shape[self.channel_axis], Rel.EQ, self.name)
         validator.check_integer("global_step rank", len(global_step_shape), 1, Rel.EQ, self.name)
         return x_shape
 
@@ -368,17 +372,17 @@ class CorrectionMul(PrimitiveWithInfer):
         >>> running_std = Tensor(np.array([2, 1.2, 0.5]), mindspore.float32)
         >>> out = correction_mul(input_x, batch_std, running_std)
     """
-    channel = 0
 
     @prim_attr_register
-    def __init__(self):
+    def __init__(self, channel_axis=0):
         """init correction mul layer"""
+        self.channel_axis = channel_axis
         self.init_prim_io_names(inputs=['x', 'batch_std', 'running_std'],
                                 outputs=['out'])
 
     def infer_shape(self, x_shape, batch_std_shape, running_std_shape):
         validator.check("batch_std shape", batch_std_shape, "running_std shape", running_std_shape, Rel.EQ, self.name)
-        validator.check("batch_std_shape[0]", batch_std_shape[0], "x_shape channel size", x_shape[self.channel],
+        validator.check("batch_std_shape[0]", batch_std_shape[0], "x_shape channel size", x_shape[self.channel_axis],
                         Rel.EQ, self.name)
         return x_shape
 
@@ -400,20 +404,20 @@ class CorrectionMulGrad(PrimitiveWithInfer):
         >>> running_std = Tensor(np.array([1.2, 0.1, 0.7, 2.3]).reshape(2, 1, 2), mindspore.float32)
         >>> result = correction_mul_grad(dout, input_x, gamma, running_std)
     """
-    channel = 0
 
     @prim_attr_register
-    def __init__(self):
+    def __init__(self, channel_axis=0):
         """init correction mul layer"""
+        self.channel_axis = channel_axis
         self.init_prim_io_names(inputs=['dout', 'x', 'gamma', 'running_std'],
                                 outputs=['dx', 'd_gamma'])
 
     def infer_shape(self, dout_shape, x_shape, gamma_shape, running_std_shape):
         validator.check("dout shape", dout_shape, "x_shape x", x_shape, Rel.EQ, self.name)
-        validator.check("gamma_shape[0]", gamma_shape[0], "dout channel size", dout_shape[self.channel],
+        validator.check("gamma_shape[0]", gamma_shape[0], "dout channel size", dout_shape[self.channel_axis],
-                        Rel.EQ, self.name)
-        validator.check("running_std_shape[0]", running_std_shape[0], "dout channel size", dout_shape[self.channel],
                         Rel.EQ, self.name)
+        validator.check("running_std_shape[0]", running_std_shape[0],
+                        "dout channel size", dout_shape[self.channel_axis], Rel.EQ, self.name)
         return x_shape, gamma_shape
 
     def infer_dtype(self, dout_type, x_type, gamma_type, running_std_type):
@@ -454,7 +458,7 @@ class BatchNormFold2(PrimitiveWithInfer):
         >>> result = batch_norm_fold2(input_x, beta, gamma, batch_std, batch_mean,
         >>>                           running_std, running_mean, global_step)
     """
-    channel = 1
+    channel_axis = 1
 
     @prim_attr_register
     def __init__(self, freeze_bn=0):
@@ -471,7 +475,7 @@ class BatchNormFold2(PrimitiveWithInfer):
         validator.check("batch_std shape", batch_std_shape, "beta shape", beta_shape, Rel.EQ, self.name)
         validator.check("batch_std shape", batch_std_shape, "running_mean shape", running_mean_shape, Rel.EQ, self.name)
         validator.check("batch_std shape", batch_std_shape, "batch_mean shape", gamma_shape, Rel.EQ, self.name)
-        validator.check("batch_std_shape[0]", batch_std_shape[0], "x_shape channel size", x_shape[self.channel],
+        validator.check("batch_std_shape[0]", batch_std_shape[0], "x_shape channel size", x_shape[self.channel_axis],
                         Rel.EQ, self.name)
         validator.check_integer("global_step rank", len(global_step_shape), 1, Rel.EQ, self.name)
         return x_shape
@@ -501,7 +505,7 @@ class BatchNormFold2Grad(PrimitiveWithInfer):
         >>> global_step = Tensor(np.array([-2]), mindspore.int32)
         >>> result = bnf2_grad(dout, input_x, gamma, batch_std, batch_mean, running_std, running_mean, global_step)
     """
-    channel = 1
+    channel_axis = 1
 
     @prim_attr_register
     def __init__(self, freeze_bn=0):
@@ -519,7 +523,7 @@ class BatchNormFold2Grad(PrimitiveWithInfer):
         validator.check("batch_std shape", batch_std_shape, "running_std shape", running_std_shape, Rel.EQ, self.name)
         validator.check("batch_std shape", batch_std_shape, "running_mean shape", running_mean_shape, Rel.EQ, self.name)
         validator.check("batch_std shape", batch_std_shape, "gamma shape", gamma_shape, Rel.EQ, self.name)
-        validator.check("batch_std size", batch_std_shape[0], "dout channel size", dout_shape[self.channel],
+        validator.check("batch_std size", batch_std_shape[0], "dout channel size", dout_shape[self.channel_axis],
                         Rel.EQ, self.name)
         validator.check_integer("global_step rank", len(global_step_shape), 1, Rel.EQ, self.name)
         return gamma_shape, gamma_shape, gamma_shape, gamma_shape, x_shape
@@ -542,3 +546,259 @@ class BatchNormFold2Grad(PrimitiveWithInfer):
         validator.check_tensor_type_same(args, (mstype.float16, mstype.float32), self.name)
         validator.check_tensor_type_same({"global_step": global_step_type}, (mstype.int32,), self.name)
         return gamma_type, gamma_type, gamma_type, gamma_type, gamma_type
+
+
+class BatchNormFoldD(PrimitiveWithInfer):
+    """Performs grad of _BatchNormFold operation."""
+
+    @prim_attr_register
+    def __init__(self, momentum=0.9, epsilon=1e-5, is_training=True, freeze_bn=0):
+        """init _BatchNormFold layer"""
+        from mindspore.ops._op_impl._custom_op import batchnorm_fold
+        self.momentum = validator.check_number_range('momentum', momentum, 0, 1, Rel.INC_BOTH, self.name)
+        self.epsilon = validator.check_float_positive('epsilon', epsilon, self.name)
+        self.is_training = validator.check_value_type('is_training', is_training, (bool,), self.name)
+        self.freeze_bn = validator.check_value_type('freeze_bn', freeze_bn, (int,), self.name)
+        self.data_format = "NCHW"
+        self.init_prim_io_names(inputs=['x', 'x_sum', 'x_square_sum', 'mean', 'variance'],
+                                outputs=['batch_mean', 'batch_std', 'running_mean', 'running_std',
+                                         'mean_updated', 'variance_updated'])
+
+    def infer_shape(self, x_shape, x_sum_shape, x_square_sum_shape, mean_shape, variance_shape):
+        validator.check("mean shape", mean_shape, "gamma_shape", variance_shape, Rel.EQ, self.name)
+        validator.check("mean_shape[0]", mean_shape[0], "input channel", x_shape[1], Rel.EQ, self.name)
+        return x_shape, mean_shape, mean_shape, mean_shape, mean_shape, mean_shape, mean_shape
+
+    def infer_dtype(self, x_type, x_sum_type, x_square_sum_type, mean_type, variance_type):
+        validator.check("input type", x_type, "mean type", mean_type)
+        validator.check("input type", x_type, "variance type", variance_type)
+        args = {"x": x_type, "mean": mean_type, "variance": variance_type}
+        validator.check_tensor_type_same(args, (mstype.float16, mstype.float32), self.name)
+        return x_type, x_type, x_type, x_type, x_type, x_type, x_type
+
+
+class BatchNormFoldGradD(PrimitiveWithInfer):
+    """Performs grad of _BatchNormFoldGrad operation."""
+
+    @prim_attr_register
+    def __init__(self, epsilon=1e-5, is_training=True, freeze_bn=0):
+        """init _BatchNormFoldGrad layer"""
+        from mindspore.ops._op_impl._custom_op import batchnorm_fold_grad
+        self.epsilon = validator.check_float_positive('epsilon', epsilon, self.name)
+        self.is_training = validator.check_value_type('is_training', is_training, (bool,), self.name)
+        self.freeze_bn = validator.check_value_type('freeze_bn', freeze_bn, (int,), self.name)
+        self.init_prim_io_names(inputs=['d_batch_mean', 'd_batch_std', 'x', 'batch_mean', 'batch_std'],
+                                outputs=['dx'])
+
+    def infer_shape(self, d_batch_mean_shape, d_batch_std_shape, x_shape, batch_mean_shape, batch_std_shape):
+        validator.check("d_batch_mean shape", d_batch_mean_shape, "d_batch_std shape", d_batch_std_shape)
+        validator.check("d_batch_mean shape", d_batch_mean_shape, "batch_mean shape", batch_mean_shape)
+        validator.check("d_batch_mean shape", d_batch_mean_shape, "batch_std shape", batch_std_shape)
+        validator.check("x_shape shape", d_batch_mean_shape[0], "input channel", x_shape[1])
+        return x_shape
+
+    def infer_dtype(self, d_batch_mean_type, d_batch_std_type, x_type, batch_mean_type, batch_std_type):
+        validator.check("input type", x_type, "d_batch_mean type", d_batch_mean_type)
+        validator.check("input type", x_type, "d_batch_std type", d_batch_std_type)
+        validator.check("input type", x_type, "batch_mean type", batch_mean_type)
+        validator.check("input type", x_type, "batch_std type", batch_std_type)
+        args = {"input type": x_type}
+        validator.check_tensor_type_same(args, (mstype.float16, mstype.float32), self.name)
+        return x_type
+
+
+class BNTrainingReduce(PrimitiveWithInfer):
+    """
+    reduce sum at axis [0, 2, 3].
+
+    Inputs:
+        - **x** (Tensor)  - Tensor of shape :math:`(N, C)`.
+
+    Outputs:
+        - **x_sum** (Tensor) - Tensor has the same shape as x.
+        - **x_square_sum** (Tensor) - Tensor has the same shape as x.
+
+    """
+
+    @prim_attr_register
+    def __init__(self):
+        """init _BNTrainingReduce layer"""
+        self.init_prim_io_names(inputs=['x'],
+                                outputs=['x_sum', 'x_square_sum'])
+
+    def infer_shape(self, x_shape):
+        return [x_shape[1]], [x_shape[1]]
+
+    def infer_dtype(self, x_type):
+        return x_type, x_type
+
+
+class BatchNormFold2_D(PrimitiveWithInfer):
+    """
+    Scale the bias with a correction factor to the long term statistics
+    prior to quantization. This ensures that there is no jitter in the quantized bias
+    due to batch to batch variation.
+
+    Inputs:
+        - **x** (Tensor)  - Tensor of shape :math:`(N, C)`.
+        - **beta** (Tensor) - Tensor of shape :math:`(C,)`.
+        - **gamma** (Tensor) - Tensor of shape :math:`(C,)`.
+        - **batch_std** (Tensor) - Tensor of shape :math:`(C,)`.
+        - **batch_mean** (Tensor) - Tensor of shape :math:`(C,)`.
+        - **running_std** (Tensor) - Tensor of shape :math:`(C,)`.
+        - **running_mean** (Tensor) - Tensor of shape :math:`(C,)`.
+        - **global_step** (Tensor) - Tensor to record current global step.
+
+    Outputs:
+        - **y** (Tensor) - Tensor has the same shape as x.
+
+    """
+    channel_axis = 1
+
+    @prim_attr_register
+    def __init__(self, freeze_bn=0):
+        """init conv2d fold layer"""
+        from mindspore.ops._op_impl._custom_op import batchnorm_fold2
+        self.init_prim_io_names(inputs=['x', 'beta', 'gamma', 'batch_std', 'batch_mean', 'running_std'],
+                                outputs=['y'])
+
+    def infer_shape(self, x_shape, beta_shape, gamma_shape, batch_std_shape, running_std_shape, batch_mean_shape):
+        validator.check("batch_std shape", batch_std_shape, "running_std shape", running_std_shape, Rel.EQ, self.name)
+        validator.check("batch_std shape", batch_std_shape, "batch_mean shape", batch_mean_shape, Rel.EQ, self.name)
+        validator.check("batch_std shape", batch_std_shape, "beta shape", beta_shape, Rel.EQ, self.name)
+        validator.check("batch_std shape", batch_std_shape, "batch_mean shape", gamma_shape, Rel.EQ, self.name)
+        validator.check("batch_std_shape[0]", batch_std_shape[0], "x_shape channel size", x_shape[self.channel_axis],
+                        Rel.EQ, self.name)
+        return x_shape
+
+    def infer_dtype(self, x_type, beta_type, gamma_type, batch_std_type, running_std_type, batch_mean_type):
+        args = {"batch_std": batch_std_type, "running_std": running_std_type, "batch_mean": batch_mean_type,
+                "beta": beta_type, "gamma": gamma_type, "x": x_type}
+        validator.check_tensor_type_same(args, (mstype.float16, mstype.float32), self.name)
+        return x_type
+
+
+class BatchNormFold2GradD(PrimitiveWithInfer):
+    """Performs grad of CorrectionAddGrad operation."""
+    channel_axis = 1
+
+    @prim_attr_register
+    def __init__(self, freeze_bn=False):
+        """init MulFold layer"""
+        from mindspore.ops._op_impl._custom_op import batchnorm_fold2_grad
+        self.freeze_bn = freeze_bn
+        self.init_prim_io_names(
+            inputs=['dout', 'dout_reduce', 'dout_x_reduce', 'gamma', 'batch_std', 'batch_mean', 'running_std'],
+            outputs=['d_batch_std', 'd_batch_mean', 'd_gamma', 'dx'])
+
+    def infer_shape(self, dout_shape, dout_reduce_shape, dout_x_reduce_shape, gamma_shape, batch_std_shape,
+                    batch_mean_shape, running_std_shape):
+        validator.check("batch_std shape", batch_std_shape, "batch_mean shape", batch_mean_shape, Rel.EQ, self.name)
+        validator.check("batch_std shape", batch_std_shape, "running_std shape", running_std_shape, Rel.EQ, self.name)
+        validator.check("batch_std shape", batch_std_shape, "gamma shape", gamma_shape, Rel.EQ, self.name)
+        validator.check("batch_std size", batch_std_shape[0], "dout channel size", dout_shape[self.channel_axis],
+                        Rel.EQ, self.name)
+        return gamma_shape, gamma_shape, gamma_shape, dout_shape
+
+    def infer_dtype(self, dout_type, dout_reduce_type, dout_x_reduce_type, gamma_type, batch_std_type,
+                    batch_mean_type, running_std_type):
+        validator.check("batch_std type", batch_std_type,
+                        "batch_mean type", batch_mean_type)
+        validator.check("batch_std type", batch_std_type,
+                        "gamma type", gamma_type)
+        validator.check("batch_std type", batch_std_type,
+                        "running_std type", running_std_type)
+        validator.check("batch_std_type", batch_std_type,
+                        "dout type", dout_type)
+        args = {"batch_std": batch_std_type, "batch_mean": batch_mean_type, "gamma": gamma_type,
+                "running_std": running_std_type, "dout": dout_type}
+        validator.check_tensor_type_same(args, (mstype.float16, mstype.float32), self.name)
+        return gamma_type, gamma_type, gamma_type, gamma_type
+
+
+class BatchNormFold2GradReduce(PrimitiveWithInfer):
+    """Performs grad of CorrectionAddGrad operation."""
+    channel_axis = 1
+
+    @prim_attr_register
+    def __init__(self, freeze_bn=False):
+        """init MulFold layer"""
+        from mindspore.ops._op_impl._custom_op import batchnorm_fold2_grad_reduce
+        self.freeze_bn = freeze_bn
+        self.init_prim_io_names(inputs=['dout', 'x'],
+                                outputs=['dout_reduce', 'dout_x_reduce'])
+
+    def infer_shape(self, dout_shape, x_shape):
+        validator.check("dout shape", dout_shape, "x shape", x_shape, Rel.EQ, self.name)
+        return (dout_shape[self.channel_axis],), (dout_shape[self.channel_axis],)
+
+    def infer_dtype(self, dout_type, x_type):
+        validator.check("dout type", dout_type, "x type", x_type)
+        return dout_type, dout_type
+
+
+class FakeQuantWithMinMaxUpdate(PrimitiveWithInfer):
+    r"""
+    Simulate the quantize and dequantize operations in training time.
+
+    Args:
+        num_bits (int) : Number bits for aware quantilization. Default: 8.
+        ema (bool): Use EMA algorithm update value min and max. Default: False.
+        ema_decay (int) : EMA algorithm decay parameter. Default: 0.999.
+        quant_delay (int): Quantilization delay parameter. Before delay step in training time not update
+            simulate aware quantize funcion. After delay step in training time begin simulate the aware
+            quantize funcion. Default: 0.
+        symmetric (bool): Quantization algorithm use symmetric or not. Default: False.
+        narrow_range (bool): Quantization algorithm use narrow range or not. Default: False.
+        training (bool): Training the network or not. Default: True.
+
+    Inputs:
+        - **x** (Tensor) : float32 Tensor representing the shape of the output tensor.
+        - **min** (Tensor) : Value of the min range of the input data x.
+        - **max** (Tensor) : Value of the max range of the input data x.
+
+    Outputs:
+        - Tensor: Simulate quantize tensor of x.
+
+    Examples:
+        >>> input_tensor = Tensor(np.random.rand(3, 16, 5, 5), mstype.float32)
+        >>> min_tensor = Tensor(np.array([-6]), mstype.float32)
+        >>> max_tensor = Tensor(np.array([6]), mstype.float32)
+        >>> output_tensor = P.FakeQuantWithMinMax(num_bits=8)(input_tensor, min_tensor, max_tensor)
+    """
+    support_quant_bit = [4, 7, 8]
+
+    @prim_attr_register
+    def __init__(self, num_bits=8, ema=False, ema_decay=0.999, quant_delay=0, symmetric=False, narrow_range=False,
+                 training=True):
+        """init FakeQuantWithMinMax OP"""
+        from mindspore.ops._op_impl._custom_op import correction_mul, correction_mul_grad
+        from mindspore.ops._op_impl._custom_op import fake_quant_with_min_max, fake_quant_with_min_max_grad
+        from mindspore.ops._op_impl._custom_op import fake_quant_with_min_max_update
+        if num_bits not in self.support_quant_bit:
+            raise ValueError(f"For '{self.name}' attr \'num_bits\' is not support.")
+        if ema and not ema_decay:
+            raise ValueError(f"For '{self.name}' attr \'ema\' and \'ema_decay\' should set together.")
+
+        self.ema = validator.check_value_type('ema', ema, (bool,), self.name)
+        self.symmetric = validator.check_value_type('symmetric', symmetric, (bool,), self.name)
+        self.narrow_range = validator.check_value_type('narrow_range', narrow_range, (bool,), self.name)
+        self.training = validator.check_value_type('training', training, (bool,), self.name)
+        self.ema_decay = validator.check_number_range('ema_decay', ema_decay, 0, 1, Rel.INC_BOTH, self.name)
+        self.num_bits = validator.check_integer('num_bits', num_bits, 0, Rel.GT, self.name)
+        self.quant_delay = validator.check_value_type('quant_delay', quant_delay, (int,), self.name)
+        self.init_prim_io_names(inputs=['x', 'min', 'max'],
+                                outputs=['min_up', 'max_up'])
+
+    def infer_shape(self, x_shape, min_shape, max_shape):
+        validator.check_integer("x rank", len(x_shape), 1, Rel.GT, self.name)
+        validator.check("min shape", min_shape, "max shape", max_shape, Rel.EQ, self.name)
+        validator.check_integer("min rank", len(min_shape), 1, Rel.EQ, self.name)
+        return min_shape, max_shape
+
+    def infer_dtype(self, x_type, min_type, max_type):
+        valid_types = (mstype.float16, mstype.float32)
+        validator.check_tensor_type_same({"x": x_type}, valid_types, self.name)
+        validator.check_tensor_type_same({"min": min_type}, valid_types, self.name)
+        validator.check_tensor_type_same({"max": max_type}, valid_types, self.name)
+        return min_type, max_type

tests/ut/python/train/quant/test_quant.py

@@ -22,7 +22,7 @@ from mindspore import nn
 from mindspore.nn.layer import combined
 from mindspore.train.quant import quant as qat
 
-context.set_context(mode=context.GRAPH_MODE)
+context.set_context(mode=context.GRAPH_MODE, device_target="GPU")
 
 
 class LeNet5(nn.Cell):
@@ -64,7 +64,7 @@ class LeNet5(nn.Cell):
         x = self.fc3(x)
         return x
 
-
+"""
 def test_qat_lenet():
     net = LeNet5()
     net = qat.convert_quant_network(
@@ -92,3 +92,4 @@ def test_qat_mobile_train():
     net = nn.WithLossCell(net, loss)
     net = nn.TrainOneStepCell(net, optimizer)
     net(img, label)
+"""