test_mkldnn.py

# Owner(s): ["module: mkldnn"]

import copy
import itertools
import functools
import unittest
from contextlib import nullcontext

try:
    import torchvision
    HAS_TORCHVISION = True
except ImportError:
    HAS_TORCHVISION = False

skipIfNoTorchVision = unittest.skipIf(not HAS_TORCHVISION, "no torchvision")

import torch
import torch.nn.functional as F
import torch.jit
import torch.backends.mkldnn
from torch.utils import mkldnn as mkldnn_utils
from torch.testing._internal.common_utils import TestCase, \
    run_tests, TemporaryFileName, gradcheck, gradgradcheck, IS_WINDOWS, \
    skipIfTorchDynamo, xfailIfTorchDynamo
from torch.testing._internal.common_device_type import (
    instantiate_device_type_tests,
    dtypes,
)

# batched grad doesn't support mkldnn
gradcheck = functools.partial(gradcheck, check_batched_grad=False)
gradgradcheck = functools.partial(gradgradcheck, check_batched_grad=False)


types = [torch.float, torch.bfloat16, torch.half]

# Comment the line below to find out the CI machines having MKL-DNN build disabled
@unittest.skipIf(not torch.backends.mkldnn.is_available(), "MKL-DNN build is disabled")
class TestMkldnn(TestCase):
    def test_conversion(self):
        for cpu_tensor in [torch.randn((1, 2, 3, 4),
                                       dtype=torch.float, device=torch.device('cpu')),
                           torch.randn((1, 2, 3, 4, 5),
                                       dtype=torch.float, device=torch.device('cpu'))[:, :, :, :, 1]]:
            cpu_tensor.requires_grad_()
            convert_dtypes = {torch.half: [torch.half, torch.float],
                              torch.bfloat16: [torch.bfloat16, torch.float],
                              torch.float: [torch.bfloat16, torch.half]}
            # float/bfloat16/half cpu tensor to mkldnn tensortensor.
            for dtype1 in types:
                mkldnn_tensor = cpu_tensor.to_mkldnn(dtype1)
                self.assertEqual(mkldnn_tensor.dtype, dtype1)
                cpu_tensor_1 = mkldnn_tensor.to_dense()
                # not given dtype for to_dense, mkldnn tensor has same dtype with cpu tensor
                self.assertEqual(mkldnn_tensor.dtype, cpu_tensor_1.dtype)
                # mkldnn float/bfloat tensor to cpu float or bfloat tensor
                for dtype2 in convert_dtypes[dtype1]:
                    cpu_tensor_2 = mkldnn_tensor.to_dense(dtype2)
                    self.assertEqual(cpu_tensor_2.dtype, dtype2)
                    atol = 1e-5 if dtype1 == torch.float and dtype2 == torch.float else 1e-2
                    self.assertEqual(cpu_tensor, cpu_tensor_2.float(), atol=atol, rtol=0)

                self.assertEqual(mkldnn_tensor.device, torch.device('cpu'))
                self.assertEqual(mkldnn_tensor.size(), torch.Size([1, 2, 3, 4]))
                self.assertEqual(mkldnn_tensor.numel(), cpu_tensor.numel())
                if dtype1 == torch.float:
                    self.assertEqual(mkldnn_tensor.element_size(), cpu_tensor.element_size())
                else:
                    self.assertEqual(mkldnn_tensor.element_size(), cpu_tensor.element_size() / 2)
                self.assertRaisesRegex(RuntimeError,
                                       "Cannot access data pointer of Tensor that doesn't have storage",
                                       lambda: mkldnn_tensor.data_ptr() != 0)

            # bfloat cpu tensor to mkldnn float tensor or bfloat tensor.
            for orig_dtype in [torch.half, torch.bfloat16]:
                cpu_tensor_lower = cpu_tensor.to(dtype=orig_dtype)
                for dtype1 in convert_dtypes[orig_dtype]:
                    mkldnn_tensor = cpu_tensor_lower.to_mkldnn(dtype1)
                    self.assertEqual(mkldnn_tensor.dtype, dtype1)
                    cpu_tensor_1 = mkldnn_tensor.to_dense()
                    # not given dtype for to_dense, mkldnn tensor has same dtype with cpu tensor
                    self.assertEqual(mkldnn_tensor.dtype, cpu_tensor_1.dtype)
                    # mkldnn float/bfloat/half tensor to cpu float/bfloat/half tensor
                    for dtype2 in convert_dtypes[cpu_tensor_lower.dtype]:
                        cpu_tensor_2 = mkldnn_tensor.to_dense(dtype2)
                        self.assertEqual(cpu_tensor_2.dtype, dtype2)
                        self.assertEqual(cpu_tensor_lower,
                                         cpu_tensor_2.to(dtype=cpu_tensor_lower.dtype), atol=1e-5, rtol=0)

                    self.assertEqual(mkldnn_tensor.device, torch.device('cpu'))
                    self.assertEqual(mkldnn_tensor.size(), torch.Size([1, 2, 3, 4]))
                    self.assertEqual(mkldnn_tensor.numel(), cpu_tensor.numel())
                    if dtype1 in [torch.bfloat16, torch.half]:
                        self.assertEqual(mkldnn_tensor.element_size(), cpu_tensor_lower.element_size())
                    else:
                        self.assertEqual(mkldnn_tensor.element_size(), cpu_tensor_lower.element_size() * 2)
                    self.assertRaisesRegex(RuntimeError,
                                           "Cannot access data pointer of Tensor that doesn't have storage",
                                           lambda: mkldnn_tensor.data_ptr() != 0)

    def test_conversion_byte_char(self):
        int8_types = [torch.int8, torch.uint8]
        for int8_type in int8_types:
            low = -100 if int8_type is torch.int8 else 0
            high = 100
            for cpu_tensor in [torch.randint(
                               low=low,
                               high=high,
                               size=(1, 2, 3, 4),
                               dtype=torch.int64,
                               device=torch.device('cpu')),
                               torch.randint(
                               low=low,
                               high=high,
                               size=(1, 2, 3, 4, 5),
                               dtype=torch.int64,
                               device=torch.device('cpu'))[:, :, :, :, :]]:

                cpu_tensor = cpu_tensor.to(dtype=int8_type)
                mkldnn_tensor = cpu_tensor.to_mkldnn(int8_type)
                self.assertEqual(mkldnn_tensor.dtype, int8_type)
                cpu_tensor_1 = mkldnn_tensor.to_dense()
                self.assertEqual(mkldnn_tensor.dtype, cpu_tensor_1.dtype)
                self.assertEqual(cpu_tensor, cpu_tensor_1)
                self.assertEqual(mkldnn_tensor.device, torch.device('cpu'))
                self.assertEqual(mkldnn_tensor.size(), cpu_tensor.size())
                self.assertEqual(mkldnn_tensor.numel(), cpu_tensor.numel())
                self.assertEqual(mkldnn_tensor.element_size(), cpu_tensor.element_size())
                self.assertRaisesRegex(RuntimeError,
                                       "Cannot access data pointer of Tensor that doesn't have storage",
                                       lambda: mkldnn_tensor.data_ptr() != 0)

    def test_copy(self):
        x = torch.randn(4, 5, dtype=torch.float32)
        mkldnn_x = x.to_mkldnn()
        mkldnn_y = torch.randn(4, 5, dtype=torch.float32).to_mkldnn()
        mkldnn_z = torch.randn(4, 10, dtype=torch.float32).to_mkldnn()
        mkldnn_y.copy_(mkldnn_x)
        self.assertEqual(x, mkldnn_y.to_dense())
        self.assertRaisesRegex(RuntimeError,
                               "copy_mkldnn_: only support same size tensor.",
                               lambda: mkldnn_z.copy_(mkldnn_x))
        self.assertRaisesRegex(RuntimeError,
                               "copy_mkldnn_: between mkldnn layout and dense Tensors is not implemented! "
                               "Found self type = torch.FloatTensor and src type = Mkldnntorch.FloatTensor",
                               lambda: x.copy_(mkldnn_x))
        self.assertRaisesRegex(RuntimeError,
                               "copy_mkldnn_: between mkldnn layout and dense Tensors is not implemented! "
                               "Found self type = Mkldnntorch.FloatTensor and src type = torch.FloatTensor",
                               lambda: mkldnn_x.copy_(x))

    def test_unsupported(self):
        # unsupported types and unsupported types with gpu
        for dtype in [torch.double, torch.uint8, torch.int8,
                      torch.short, torch.int, torch.long]:
            with self.assertRaises(RuntimeError):
                torch.randn(1, 2, 3, 4, dtype=dtype, device=torch.device('cpu')).to_mkldnn()
            if torch.cuda.is_available():
                with self.assertRaises(RuntimeError):
                    torch.randn(1, 2, 3, 4, dtype=dtype, device=torch.device('cuda')).to_mkldnn()
        # supported type with gpu
        if torch.cuda.is_available():
            with self.assertRaises(RuntimeError):
                torch.randn(1, 2, 3, 4, dtype=torch.float, device=torch.device('cuda')).to_mkldnn()
        # some factory functions
        for creator in [torch.ones, torch.randn, torch.rand]:
            with self.assertRaises(RuntimeError):
                creator(1, 2, 3, 4, dtype=torch.float, device=torch.device('cpu'), layout=torch._mkldnn)

    def test_mkldnn_conv_shapecheck(self):
        input = torch.full((1, 1, 1, 24,), 1, dtype=torch.float32)
        w1 = torch.full((1, 1, 1, 24,), 1, dtype=torch.float32)
        b1 = torch.full((1,), 1, dtype=torch.float32)
        w2 = torch.full((1, 1, 2, 24,), 1, dtype=torch.float32)
        b2 = torch.full((2,), 1, dtype=torch.float32)
        options = zip([-1, 0, 0, 0, 0, 0, 0],  # padding
                      [1, 0, 1, 1, 1, 1, 1],  # stride
                      [1, 1, 0, 1, 1, 1, 1],  # dilation
                      [1, 1, 1, 0, 2, 1, 1],  # groups
                      [w1, w1, w1, w1, w1, w1, w2],  # weight
                      [b1, b1, b1, b1, b1, b2, b1])  # bias
        for pad, st, dil, gr, w, b in options:
            with self.assertRaises(RuntimeError) as _:
                torch.mkldnn_convolution(input, w, b, [pad] * 2, [st] * 2, [dil] * 2, gr)

    def test_autograd_to_mkldnn(self):
        # MKLDNN only supports float32
        root = torch.randn(4, 5, dtype=torch.float32, requires_grad=True)

        def func(root):
            return root.to_mkldnn().to_dense()

        # because MKLDNN only supports float32, we need to lessen the precision.
        # these numbers are just empirical results that seem to work.
        self.assertWarnsRegex(UserWarning,
                              'double precision floating point',
                              lambda: gradcheck(func, [root], atol=4e-2, rtol=1e-2))
        self.assertWarnsRegex(UserWarning,
                              'double precision floating point',
                              lambda: gradgradcheck(func, [root], atol=4e-2, rtol=1e-2))

    def test_autograd_from_mkldnn(self):
        # MKLDNN only supports float32
        root = torch.randn(4, 5, dtype=torch.float32).to_mkldnn().requires_grad_()

        def func(root):
            return root.to_dense()

        # because MKLDNN only supports float32, we need to lessen the precision.
        # these numbers are just empirical results that seem to work.
        self.assertWarnsRegex(UserWarning,
                              'double precision floating point',
                              lambda: gradcheck(func, [root], atol=4e-2, rtol=1e-2))

    def test_detach(self):
        root = torch.randn(4, 5, dtype=torch.float32).to_mkldnn().requires_grad_()

        detach = root.detach()
        self.assertEqual((4, 5), detach.size())
        self.assertFalse(detach.requires_grad)
        self.assertTrue(root.requires_grad)

        detach_ = root.detach_()
        self.assertEqual((4, 5), detach_.size())
        self.assertFalse(detach_.requires_grad)
        self.assertFalse(root.requires_grad)

    def test_repr(self):
        self.assertTrue("layout=torch._mkldnn" in str(torch.randn((1, 2, 3, 4),
                                                                  dtype=torch.float, device=torch.device('cpu')).to_mkldnn()))

    def _test_conv_base(self, dim):
        conv_module = {1: torch.nn.Conv1d, 2: torch.nn.Conv2d, 3: torch.nn.Conv3d}
        input_shapes = {1: (224,), 2: (224, 224), 3: (55, 55, 55)}
        options = itertools.product([True, False], [True, False], [1, 2], [1, 4])
        for train, bias, dilation, groups in options:
            N = torch.randint(3, 10, (1,)).item()
            M = torch.randint(1, 3, (1,)).item() * groups
            C = torch.randint(1, 3, (1,)).item() * groups
            x_shape = (N, C) + input_shapes[dim]
            x = torch.randn(x_shape, dtype=torch.float32)
            conv = conv_module[dim](in_channels=C,
                                    out_channels=M,
                                    kernel_size=3,
                                    stride=2,
                                    padding=1,
                                    dilation=dilation,
                                    bias=bias,
                                    groups=groups).float()
            x1 = x.clone()
            x2 = x.clone().to_mkldnn()
            if not train:
                mkldnn_conv = mkldnn_utils.to_mkldnn(copy.deepcopy(conv))
            elif train and dim != 1:
                # TODO: enable conv1d training.
                x1.requires_grad_()
                x2.requires_grad_()
                mkldnn_conv = copy.deepcopy(conv)
            with torch.backends.mkldnn.flags(enabled=False):
                y_aten = conv(x1)
                if train and dim != 1:
                    loss1 = y_aten.sum()
                    loss1.backward()
            if not train or (train and dim != 1):
                y_mkldnn = mkldnn_conv(x2).to_dense()
                self.assertEqual(y_aten, y_mkldnn)
            if not train:
                self._test_serialization(mkldnn_conv, (x.to_mkldnn(),))
                self._test_tracing(mkldnn_conv, (x.to_mkldnn(),))
            elif dim != 1:
                loss2 = y_mkldnn.sum()
                loss2.backward()
                self.assertTrue(x2.grad.is_mkldnn)
                self.assertEqual(x1.grad, x2.grad.to_dense())
                self.assertEqual(conv.weight.grad,
                                 mkldnn_conv.weight.grad,
                                 atol=1e-3,
                                 rtol=1e-3)
                if bias:
                    self.assertEqual(conv.bias.grad, mkldnn_conv.bias.grad)

    def test_conv1d(self):
        self._test_conv_base(dim=1)

    def test_conv2d(self):
        self._test_conv_base(dim=2)

    def test_conv3d(self):
        self._test_conv_base(dim=3)

    def _test_conv_deconv_lower_precision_base(self, dim, conv_module, dtype):
        input_shapes = {1: (224,), 2: (224, 224), 3: (55, 55, 55)}
        options = itertools.product([True, False], [1, 2], [1, 4])
        for bias, dilation, groups in options:
            N = torch.randint(1, 3, (1,)).item()
            M = torch.randint(1, 3, (1,)).item() * groups
            C = torch.randint(1, 3, (1,)).item() * groups
            x_shape = (N, C) + input_shapes[dim]
            x = torch.randn(x_shape, dtype=torch.float32)
            # TODO: remove this when group depthwise is supported:
            if conv_module in [torch.nn.ConvTranspose1d, torch.nn.ConvTranspose2d,
                               torch.nn.ConvTranspose3d] and groups > 1 and C == groups:
                continue
            conv = conv_module(in_channels=C,
                               out_channels=M,
                               kernel_size=3,
                               stride=2,
                               padding=1,
                               dilation=dilation,
                               bias=bias,
                               groups=groups).float()
            x_lower = x.to(dtype=dtype)
            if (dtype == torch.bfloat16 and torch.ops.mkldnn._is_mkldnn_bf16_supported()) or \
               (dtype == torch.half and torch.ops.mkldnn._is_mkldnn_fp16_supported()):
                mkldnn_conv = mkldnn_utils.to_mkldnn(copy.deepcopy(conv))
                mkldnn_conv_lower = mkldnn_utils.to_mkldnn(copy.deepcopy(conv), dtype)
                y = mkldnn_conv(x.to_mkldnn()).to_dense()
                y_lower = mkldnn_conv_lower(x_lower.to_mkldnn()).to_dense(torch.float32)
                self.assertEqual(y, y_lower, atol=1e-1, rtol=1e-3)
            else:
                msg = {
                    torch.bfloat16: r"bf16 path needs the cpu support avx_ne_convert or avx512bw, avx512vl and avx512dq",
                    torch.half: r"fp16 path needs the cpu support avx_ne_convert or avx512_fp16",
                }
                with self.assertRaisesRegex(RuntimeError, msg[dtype]):
                    mkldnn_conv_lower = mkldnn_utils.to_mkldnn(copy.deepcopy(conv), dtype)
                    y_lower = mkldnn_conv_lower(x_lower.to_mkldnn()).to_dense(torch.float32)
            # test thnn impl
            conv_lower = copy.deepcopy(conv).to(dtype=dtype)
            conv_ref = copy.deepcopy(conv_lower).float()
            with torch.backends.mkldnn.flags(enabled=False):
                x_ref = x_lower.clone().float().detach().requires_grad_()
                x_lower.requires_grad_()
                y = conv_ref(x_ref)
                y_lower = conv_lower(x_lower).float()
                self.assertEqual(y, y_lower, atol=5e-2, rtol=5e-3)

    @dtypes(torch.float16, torch.bfloat16)
    def test_conv_deconv_1d_lower_precision(self, dtype):
        self._test_conv_deconv_lower_precision_base(1, torch.nn.Conv1d, dtype=dtype)
        self._test_conv_deconv_lower_precision_base(1, torch.nn.ConvTranspose1d, dtype=dtype)

    @dtypes(torch.float16, torch.bfloat16)
    def test_conv_deconv_2d_lower_precision(self, dtype):
        self._test_conv_deconv_lower_precision_base(2, torch.nn.Conv2d, dtype=dtype)
        self._test_conv_deconv_lower_precision_base(2, torch.nn.ConvTranspose2d, dtype=dtype)

    @dtypes(torch.float16, torch.bfloat16)
    def test_conv_deconv_3d_lower_precision(self, dtype):
        self._test_conv_deconv_lower_precision_base(3, torch.nn.Conv3d, dtype=dtype)
        self._test_conv_deconv_lower_precision_base(3, torch.nn.ConvTranspose3d, dtype=dtype)

    def _test_conv_deconv_nhwc_base(self, conv_module, weight_memory_format, dtype, prec=None):
        input_shapes = {2: (55, 55), 3: (14, 14, 14)}
        options = itertools.product([True, False], [True, False], [1, 2], [1, 4])
        if conv_module in [torch.nn.Conv2d, torch.nn.ConvTranspose2d]:
            cl_format = torch.channels_last
            input_shape = input_shapes[2]
        elif conv_module in [torch.nn.Conv3d, torch.nn.ConvTranspose3d]:
            cl_format = torch.channels_last_3d
            input_shape = input_shapes[3]

        for train, bias, dilation, groups in options:
            N = torch.randint(3, 10, (1,)).item()
            M = torch.randint(1, 3, (1,)).item() * groups
            C = torch.randint(1, 3, (1,)).item() * groups
            x_shape = (N, C) + input_shape
            x = torch.randn(x_shape, dtype=dtype)

            # conv1: mkldnn conv/deconv in contiguous memory format (nchw)
            # conv2: mkldnn conv/deconv in channels last memory format (nhwc)
            conv1 = conv_module(in_channels=C,
                                out_channels=M,
                                kernel_size=3,
                                stride=2,
                                padding=1,
                                dilation=dilation,
                                bias=bias,
                                groups=groups).to(dtype=dtype)
            conv2 = copy.deepcopy(conv1).to(memory_format=weight_memory_format)
            x1 = x.clone()
            x2 = x.clone().to(memory_format=cl_format)
            if train:
                x1.requires_grad_()
                x2.requires_grad_()
            y1 = conv1(x1)
            y2 = conv2(x2)
            self.assertEqual(y1, y2, atol=prec, rtol=prec)

            if train:
                y1.sum().backward()
                y2.sum().backward()
                self.assertTrue(x2.grad.is_contiguous(memory_format=cl_format))
                self.assertEqual(conv1.weight.grad,
                                 conv2.weight.grad,
                                 atol=1e-3,
                                 rtol=1e-3)
                if bias:
                    self.assertEqual(conv1.bias.grad, conv2.bias.grad, atol=prec, rtol=prec)
                self.assertEqual(x1.grad, x2.grad, atol=prec, rtol=prec)

    def test_conv_nhwc_fp32(self):
        self._test_conv_deconv_nhwc_base(torch.nn.Conv2d, torch.contiguous_format, dtype=torch.float32)
        self._test_conv_deconv_nhwc_base(torch.nn.Conv2d, torch.channels_last, dtype=torch.float32)
        self._test_conv_deconv_nhwc_base(torch.nn.Conv3d, torch.contiguous_format, dtype=torch.float32)
        self._test_conv_deconv_nhwc_base(torch.nn.Conv3d, torch.channels_last_3d, dtype=torch.float32)

    @dtypes(torch.float16, torch.bfloat16)
    def test_conv_nhwc_lower_precision(self, dtype):
        # when torch.ops.mkldnn._is_mkldnn_bf16_supported() or torch.ops.mkldnn._is_mkldnn_fp16_supported()
        # returns false, bf16/fp16 CPU conv will fall back to thnn impl
        support_checks = {
            torch.bfloat16: torch.ops.mkldnn._is_mkldnn_bf16_supported,
            torch.float16: torch.ops.mkldnn._is_mkldnn_fp16_supported
        }
        if support_checks[dtype]():
            self._test_conv_deconv_nhwc_base(torch.nn.Conv2d, torch.contiguous_format, dtype=dtype)
            self._test_conv_deconv_nhwc_base(torch.nn.Conv2d, torch.channels_last, dtype=dtype)
            self._test_conv_deconv_nhwc_base(torch.nn.Conv3d, torch.contiguous_format, dtype=dtype)
            self._test_conv_deconv_nhwc_base(torch.nn.Conv3d, torch.channels_last_3d, dtype=dtype)

        # BF16/FP16 fallback implementations are divided into two parts im2col+gemm,
        # and the number of data type conversions in the middle is more than that of onednn's direct conv,
        # resulting in additional accuracy loss.
        precisions = {
            torch.bfloat16: 1e-2,
            torch.float16: 2e-3,
        }
        prec = precisions[dtype]
        with torch.backends.mkldnn.flags(enabled=False):
            self._test_conv_deconv_nhwc_base(torch.nn.Conv2d, torch.contiguous_format, dtype=dtype, prec=prec)
            self._test_conv_deconv_nhwc_base(torch.nn.Conv2d, torch.channels_last, dtype=dtype, prec=prec)
            self._test_conv_deconv_nhwc_base(torch.nn.Conv3d, torch.contiguous_format, dtype=dtype, prec=prec)
            self._test_conv_deconv_nhwc_base(torch.nn.Conv3d, torch.channels_last_3d, dtype=dtype, prec=prec)


    def test_conv_transpose_nhwc_fp32(self):
        self._test_conv_deconv_nhwc_base(torch.nn.ConvTranspose2d, torch.contiguous_format, dtype=torch.float32)
        self._test_conv_deconv_nhwc_base(torch.nn.ConvTranspose2d, torch.channels_last, dtype=torch.float32)
        self._test_conv_deconv_nhwc_base(torch.nn.ConvTranspose3d, torch.contiguous_format, dtype=torch.float32)
        self._test_conv_deconv_nhwc_base(torch.nn.ConvTranspose3d, torch.channels_last_3d, dtype=torch.float32)

    @dtypes(torch.float16, torch.bfloat16)
    def test_conv_transpose_nhwc_lower_precision(self, dtype):
        # when torch.ops.mkldnn._is_mkldnn_bf16_supported() or torch.ops.mkldnn._is_mkldnn_fp16_supported()
        # returns false, bf16/fp16 CPU conv will fall back to thnn impl
        support_checks = {
            torch.bfloat16: torch.ops.mkldnn._is_mkldnn_bf16_supported,
            torch.float16: torch.ops.mkldnn._is_mkldnn_fp16_supported
        }
        if support_checks[dtype]():
            self._test_conv_deconv_nhwc_base(torch.nn.ConvTranspose2d, torch.contiguous_format, dtype=dtype)
            self._test_conv_deconv_nhwc_base(torch.nn.ConvTranspose2d, torch.channels_last, dtype=dtype)
            self._test_conv_deconv_nhwc_base(torch.nn.ConvTranspose3d, torch.contiguous_format, dtype=dtype)
            self._test_conv_deconv_nhwc_base(torch.nn.ConvTranspose3d, torch.channels_last_3d, dtype=dtype)

        # BF16/FP16 fallback implementations are divided into two parts col2im+gemm,
        # and the number of data type conversions in the middle is more than that of onednn's direct conv,
        # resulting in additional accuracy loss.
        precisions = {
            torch.bfloat16: 2e-2,
            torch.float16: 3e-3,
        }
        prec = precisions[dtype]
        with torch.backends.mkldnn.flags(enabled=False):
            self._test_conv_deconv_nhwc_base(torch.nn.ConvTranspose2d, torch.contiguous_format, dtype=dtype, prec=prec)
            self._test_conv_deconv_nhwc_base(torch.nn.ConvTranspose2d, torch.channels_last, dtype=dtype, prec=prec)
            self._test_conv_deconv_nhwc_base(torch.nn.ConvTranspose3d, torch.contiguous_format, dtype=dtype, prec=prec)
            self._test_conv_deconv_nhwc_base(torch.nn.ConvTranspose3d, torch.channels_last_3d, dtype=dtype, prec=prec)

    def _test_conv_transpose_base(self, dim):
        conv_module = {
            1: torch.nn.ConvTranspose1d,
            2: torch.nn.ConvTranspose2d,
            3: torch.nn.ConvTranspose3d
        }
        input_shapes = {1: (55,), 2: (28, 28), 3: (14, 14, 14)}
        options = itertools.product([True, False], [True, False], [1, 2], [1, 4])
        for train, bias, dilation, groups in options:
            N = torch.randint(3, 10, (1,)).item()
            M = torch.randint(1, 3, (1,)).item() * groups
            C = torch.randint(1, 3, (1,)).item() * groups
            x_shape = (N, C) + input_shapes[dim]
            data = torch.randn(x_shape, dtype=torch.float32)
            # conv: mkldnn tranpose conv fp32
            # conv_ref: thnn transpose conv fp32
            conv = conv_module[dim](in_channels=C,
                                    out_channels=M,
                                    kernel_size=3,
                                    stride=1,
                                    padding=1,
                                    dilation=dilation,
                                    bias=bias,
                                    groups=groups).to(dtype=torch.float32)
            x = data.clone()
            x_ref = x.clone()
            if train:
                x.requires_grad_()
                x_ref.requires_grad_()

            conv_ref = copy.deepcopy(conv)
            with torch.backends.mkldnn.flags(enabled=False):
                y_ref = conv_ref(x_ref)
                if train:
                    y_ref.sum().backward()

            y = conv(x)
            if train:
                y.sum().backward()

            self.assertEqual(y, y_ref)
            if train:
                self.assertEqual(x.grad, x_ref.grad)
                self.assertEqual(conv.weight.grad,
                                 conv_ref.weight.grad,
                                 atol=1e-3,
                                 rtol=1e-3)
                if bias:
                    self.assertEqual(conv.bias.grad, conv_ref.bias.grad)

    def test_conv_transpose1d(self):
        self._test_conv_transpose_base(dim=1)

    def test_conv_transpose2d(self):
        self._test_conv_transpose_base(dim=2)

    def test_conv_transpose3d(self):
        self._test_conv_transpose_base(dim=3)

    def test_conv2d_legacy_jit_model(self):
        """
        MKLDNN integration used to serialize models with 5d weight for grouped
        convolutions, we'd like to preserve this behavior
        """
        g = 4
        conv2d = torch.nn.Conv2d(16, 16, 3, groups=g)
        conv2d_mkldnn = torch.utils.mkldnn.to_mkldnn(conv2d)

        # contrive legacy conv2d module with a 5-d weight
        o, i, h, w = conv2d.weight.shape
        weight_5d = conv2d.weight.reshape((g, o // g, i, h, w))
        conv2d_mkldnn.weight = weight_5d.to_mkldnn()

        x = torch.randn(1, 16, 8, 8)

        with TemporaryFileName() as fname:
            torch.jit.save(conv2d_mkldnn, fname)
            conv2d_loaded = torch.jit.load(fname)

            self.assertEqual(conv2d_mkldnn.weight.ndimension(), 5)
            self.assertEqual(conv2d_loaded.weight.ndimension(), 4)
            self.assertEqual(
                conv2d(x),
                conv2d_loaded(x.to_mkldnn()).to_dense())

    # This test is to check whether 1D conv is supported for mkldnn tensor,
    # which is exposed by Issue https://github.com/pytorch/pytorch/issues/68034.
    def test_conv1d_functional(self):
        input = torch.randn(2, 3, 10).to_mkldnn()
        weight = torch.randn(3, 3, 3).to_mkldnn()
        bias = torch.randn(3).to_mkldnn()
        output = torch.nn.functional.conv1d(input, weight, bias)
        self.assertEqual(output.size(), torch.Size([2, 3, 8]))

    def test_relu(self):
        x = torch.randn((4, 5), dtype=torch.float32) * 10
        x1 = x.clone().requires_grad_()
        x2 = x.clone().to_mkldnn().requires_grad_()
        y1 = torch.relu(x1)
        y2 = torch.relu(x2).to_dense()
        loss1 = y1.sum()
        loss2 = y2.sum()
        loss1.backward()
        loss2.backward()
        self.assertEqual(y1, y2)
        self.assertEqual(x1.grad, x2.grad.to_dense())

    def test_relu_(self):
        x = torch.randn((4, 5), dtype=torch.float32) * 10
        x1 = x.clone().requires_grad_()
        x2 = x.clone().to_mkldnn().requires_grad_()
        y1 = torch.relu_(x1.clone())
        y2 = torch.relu_(x2.clone()).to_dense()
        loss1 = y1.sum()
        loss2 = y2.sum()
        loss1.backward()
        loss2.backward()
        self.assertEqual(y1, y2)
        self.assertEqual(x1.grad, x2.grad.to_dense())

    @unittest.skipIf(IS_WINDOWS, "Limit support for bf16 path")
    def _test_relu_bf16_base(self, name):
        x = torch.randn((4, 5), dtype=torch.float32) * 10
        x_bf16 = x.bfloat16()
        fn = getattr(torch, name)
        if torch.ops.mkldnn._is_mkldnn_bf16_supported():
            y = fn(x.to_mkldnn()).to_dense()
            y_bf16 = fn(x_bf16.to_mkldnn()).to_dense(torch.float32)
            self.assertEqual(y, y_bf16, atol=1e-1, rtol=1e-3)
        else:
            msg = r"bf16 path needs the cpu support avx512bw, avx512vl and avx512dq"
            self.assertRaisesRegex(RuntimeError,
                                   msg,
                                   lambda: fn(x_bf16.to_mkldnn()))

    def test_relu_bf16(self):
        self._test_relu_bf16_base("relu")

    def test_relu_inplace_bf16(self):
        self._test_relu_bf16_base("relu_")

    def test_gelu(self):
        m = torch.nn.GELU()
        x = torch.randn((4, 5), dtype=torch.float32) * 10
        x1 = x.clone().requires_grad_()
        x2 = x.clone().to_mkldnn().requires_grad_()
        y1 = m(x1)
        y2 = m(x2).to_dense()
        loss1 = y1.sum()
        loss2 = y2.sum()
        loss1.backward()
        loss2.backward()
        self.assertEqual(y1, y2)
        self.assertEqual(x1.grad, x2.grad.to_dense())

    @unittest.skipIf(IS_WINDOWS, "Limit support for bf16 path")
    def test_gelu_bf16(self):
        m = torch.nn.GELU()
        x = torch.randn((4, 5), dtype=torch.float32) * 10
        x1 = x.clone().to_mkldnn().requires_grad_()
        x2 = x.clone().to_mkldnn(torch.bfloat16).requires_grad_()
        if torch.ops.mkldnn._is_mkldnn_bf16_supported():
            y1 = m(x1).to_dense()
            y2 = m(x2).to_dense()
            loss1 = y1.sum()
            loss2 = y2.sum()
            loss1.backward()
            loss2.backward()
            self.assertEqual(y1, y2.to(torch.float32), atol=1e-1, rtol=0)
            self.assertEqual(x1.grad.to_dense(), x2.grad.to_dense(torch.float32), atol=1e-2, rtol=0)
        else:
            msg = r"bf16 path needs the cpu support avx512bw, avx512vl and avx512dq"
            self.assertRaisesRegex(RuntimeError,
                                   msg,
                                   lambda: m(x2))

    def _test_prelu_base(self, size, num_channels):
        x = torch.randn(size, dtype=torch.float32)
        x1 = x.clone().requires_grad_()
        x2 = x.clone().to_mkldnn().requires_grad_()
        x3 = x.clone().to_mkldnn().requires_grad_()
        m1 = torch.nn.PReLU(num_channels)
        m2 = mkldnn_utils.to_mkldnn(copy.deepcopy(m1))
        m3 = copy.deepcopy(m1)
        y1 = m1(x1)
        y2 = m2(x2).to_dense()
        y3 = m3(x3).to_dense()  # Only convert data to mkldnn, weight is Aten tensor
        loss1 = y1.sum()
        loss1.backward()
        loss2 = y2.sum()
        loss2.backward()
        loss3 = y3.sum()
        loss3.backward()
        self.assertEqual(y1, y2)
        self.assertEqual(y1, y3)
        self.assertEqual(x1.grad, x2.grad.to_dense())
        self.assertEqual(x1.grad, x3.grad.to_dense())

    def test_prelu(self):
        self._test_prelu_base(torch.Size([16]), 1)
        self._test_prelu_base(torch.Size([16, 64]), 1)
        self._test_prelu_base(torch.Size([16, 64]), 64)
        self._test_prelu_base(torch.Size([16, 64, 112]), 1)
        self._test_prelu_base(torch.Size([16, 64, 112]), 64)
        self._test_prelu_base(torch.Size([16, 64, 112, 112]), 1)
        self._test_prelu_base(torch.Size([16, 64, 112, 112]), 64)
        self._test_prelu_base(torch.Size([16, 64, 112, 112, 1]), 1)
        self._test_prelu_base(torch.Size([16, 64, 112, 112, 1]), 64)

    @unittest.skipIf(IS_WINDOWS, "Limit support for bf16 path")
    def _test_prelu_bf16_base(self, size, num_channels):
        if torch.ops.mkldnn._is_mkldnn_bf16_supported():
            x = torch.randn(size, dtype=torch.float32)
            x_fp32 = x.clone().to_mkldnn().requires_grad_()
            x_bf16 = x.clone().to_mkldnn(torch.bfloat16).requires_grad_()
            m = mkldnn_utils.to_mkldnn(torch.nn.PReLU())
            m_bf16 = mkldnn_utils.to_mkldnn(torch.nn.PReLU(), torch.bfloat16)

            y = m(x_fp32).to_dense()
            y_bf16 = m_bf16(x_bf16).to_dense()
            self.assertEqual(y, y_bf16.to(torch.float32), atol=1e-1, rtol=1e-3)

            loss = y.sum()
            loss.backward()
            loss_bf16 = y_bf16.sum()
            loss_bf16.backward()
            self.assertEqual(x_fp32.grad.to_dense(), x_bf16.grad.to_dense(torch.float32))
        else:
            x_bf16 = torch.randn(size, dtype=torch.bfloat16).requires_grad_()
            m_bf16 = mkldnn_utils.to_mkldnn(torch.nn.PReLU(), torch.bfloat16)
            msg = r"bf16 path needs the cpu support avx512bw, avx512vl and avx512dq"
            self.assertRaisesRegex(RuntimeError,
                                   msg,
                                   lambda: m_bf16(x_bf16))

    def test_prelu_bf16(self):
        self._test_prelu_bf16_base(torch.Size([16]), 1)
        self._test_prelu_bf16_base(torch.Size([16, 64]), 1)
        self._test_prelu_bf16_base(torch.Size([16, 64]), 64)
        self._test_prelu_bf16_base(torch.Size([16, 64, 112]), 1)
        self._test_prelu_bf16_base(torch.Size([16, 64, 112]), 64)
        self._test_prelu_bf16_base(torch.Size([16, 64, 112, 112, 1]), 1)
        self._test_prelu_bf16_base(torch.Size([16, 64, 112, 112, 1]), 64)

    def _test_max_pool_base(self, dim, input):
        pool_module = {2: torch.nn.MaxPool2d, 3: torch.nn.MaxPool3d}
        for stride in [1, 2, 3]:
            for ceil_mode in [False, True]:
                max_pool = pool_module[dim](
                    kernel_size=3 if not ceil_mode else 7,
                    stride=stride,
                    padding=1,
                    ceil_mode=ceil_mode)

                x1 = input.clone().requires_grad_()
                x2 = input.clone().to_mkldnn().requires_grad_()
                y1 = max_pool(x1)
                y2 = max_pool(x2).to_dense()
                loss1 = y1.sum()
                loss2 = y2.sum()
                loss1.backward()
                loss2.backward()
                self.assertEqual(y1, y2)
                self.assertEqual(x1.grad, x2.grad.to_dense())

    def test_max_pool2d(self):
        N = torch.randint(3, 10, (1,)).item()
        C = torch.randint(3, 10, (1,)).item()
        for H, W in [(64, 64), (35, 39), (16, 19), [7, 8]]:
            x = torch.randn(N, C, H, W, dtype=torch.float32) * 10
            self._test_max_pool_base(dim=2, input=x)

    def test_max_pool3d(self):
        N = torch.randint(3, 10, (1,)).item()
        C = torch.randint(3, 10, (1,)).item()
        for D, H, W in [(64, 64, 64), (35, 39, 35), (16, 19, 20), [7, 8, 9]]:
            x = torch.randn(N, C, D, H, W, dtype=torch.float32) * 10
            self._test_max_pool_base(dim=3, input=x)


    @unittest.skipIf(IS_WINDOWS, "Limit support for bf16 path")
    def _test_max_pool_bf16_base(self, dim, input):
        pool_module = {2: torch.nn.MaxPool2d, 3: torch.nn.MaxPool3d}
        x_bf16 = input.bfloat16()
        for stride in [1, 2, 3]:
            for ceil_mode in [False, True]:
                max_pool = pool_module[dim](
                    kernel_size=3 if not ceil_mode else 7,
                    stride=stride,
                    padding=1,
                    ceil_mode=ceil_mode)

                if torch.ops.mkldnn._is_mkldnn_bf16_supported():
                    y = max_pool(input.to_mkldnn()).to_dense()
                    y_bf16 = max_pool(x_bf16.to_mkldnn()).to_dense(torch.float32)
                    self.assertEqual(y, y_bf16, atol=0.1, rtol=1e-3)
                else:
                    msg = f"mkldnn_max_pool{dim:d}d: bf16 path needs the cpu support avx512bw, avx512vl and avx512dq"
                    self.assertRaisesRegex(RuntimeError,
                                           msg,
                                           lambda: max_pool(x_bf16.to_mkldnn()))

    def test_max_pool2d_bf16(self):
        N = torch.randint(3, 10, (1,)).item()
        C = torch.randint(3, 10, (1,)).item()
        for H, W in [(64, 64), (35, 39), (16, 19), [7, 8]]:
            x = torch.randn(N, C, H, W, dtype=torch.float32) * 10
            self._test_max_pool_bf16_base(dim=2, input=x)

    def test_max_pool3d_bf16(self):
        N = torch.randint(3, 10, (1,)).item()
        C = torch.randint(3, 10, (1,)).item()
        for D, H, W in [(64, 64, 64), (35, 39, 35), (16, 19, 20), [7, 8, 9]]:
            x = torch.randn(N, C, D, H, W, dtype=torch.float32) * 10
            self._test_max_pool_bf16_base(dim=3, input=x)

    def test_max_pool2d_stride_none(self):
        N = torch.randint(3, 10, (1,)).item()
        C = torch.randint(3, 10, (1,)).item()

        for H, W in [(64, 64), (35, 39), (16, 19), [7, 8]]:
            x = torch.randn(N, C, H, W, dtype=torch.float32) * 10
            for ceil_mode in [False, True]:
                y1 = F.max_pool2d(
                    x,
                    kernel_size=3 if not ceil_mode else 7,
                    stride=None,
                    padding=1,
                    ceil_mode=ceil_mode)

                y2 = F.max_pool2d(
                    x.to_mkldnn(),
                    kernel_size=3 if not ceil_mode else 7,
                    stride=None,
                    padding=1,
                    ceil_mode=ceil_mode)

                self.assertEqual(y1, y2.to_dense())

    # https://github.com/pytorch/pytorch/issues/127111
    @xfailIfTorchDynamo
    def test_max_pool_unsupported(self):
        # OneDNN not support dilation max_pooling, will be avilabled in v2.0.
        N = torch.randint(3, 10, (1,)).item()
        C = torch.randint(3, 10, (1,)).item()

        # 2d dilation case
        x = torch.randn(N, C, 7, 7, dtype=torch.float32).to_mkldnn()
        max_pool2d = torch.nn.MaxPool2d(
            kernel_size=3,
            stride=3,
            padding=1,
            dilation=2)
        self.assertRaisesRegex(RuntimeError,
                               'mkldnn_max_pool2d does not support dilation case',
                               lambda: max_pool2d(x))

        # 3d dilation case
        x = torch.randn(N, C, 7, 7, 7, dtype=torch.float32).to_mkldnn()
        max_pool3d = torch.nn.MaxPool3d(
            kernel_size=3,
            stride=3,
            padding=1,
            dilation=2)
        self.assertRaisesRegex(RuntimeError,
                               'mkldnn_max_pool3d does not support dilation case',
                               lambda: max_pool3d(x))

    def _test_avg_pool_base(self, dim, input):
        avg_module = {2: torch.nn.AvgPool2d, 3: torch.nn.AvgPool3d}
        for count_include_pad in [True, False]:
            avg_pool = avg_module[dim](
                kernel_size=3,
                stride=2,
                padding=1,
                count_include_pad=count_include_pad)

            x1 = input.clone().requires_grad_()
            x2 = input.clone().to_mkldnn().requires_grad_()
            y1 = avg_pool(x1)
            y2 = avg_pool(x2).to_dense()
            loss1 = y1.sum()
            loss2 = y2.sum()
            loss1.backward()
            loss2.backward()
            self.assertEqual(y1, y2)
            self.assertEqual(x1.grad, x2.grad.to_dense())

    def test_avg_pool2d(self):
        N = torch.randint(3, 10, (1,)).item()
        C = torch.randint(3, 10, (1,)).item()
        x = torch.randn(N, C, 64, 64, dtype=torch.float32) * 10
        self._test_avg_pool_base(dim=2, input=x)

    def test_avg_pool3d(self):
        N = torch.randint(3, 10, (1,)).item()
        C = torch.randint(3, 10, (1,)).item()
        x = torch.randn(N, C, 64, 64, 64, dtype=torch.float32) * 10
        self._test_avg_pool_base(dim=3, input=x)

    @unittest.skipIf(IS_WINDOWS, "Limit support for bf16 path")
    def _test_avg_pool_bf16_base(self, dim, input):
        avg_module = {2: torch.nn.AvgPool2d, 3: torch.nn.AvgPool3d}
        x_bf16 = input.bfloat16()
        for count_include_pad in [True, False]:
            avg_pool = avg_module[dim](
                kernel_size=3,
                stride=2,
                padding=1,
                count_include_pad=count_include_pad)
            if torch.ops.mkldnn._is_mkldnn_bf16_supported():
                y = avg_pool(input.to_mkldnn()).to_dense()
                y_bf16 = avg_pool(x_bf16.to_mkldnn()).to_dense(torch.float)
                self.assertEqual(y, y_bf16, atol=1e-1, rtol=1e-3)
            else:
                msg = f"mkldnn_avg_pool{dim:d}d: bf16 path needs the cpu support avx512bw, avx512vl and avx512dq"
                self.assertRaisesRegex(RuntimeError,
                                       msg,
                                       lambda: avg_pool(x_bf16.to_mkldnn()))

    def test_avg_pool2d_bf16(self):
        N = torch.randint(3, 10, (1,)).item()
        C = torch.randint(3, 10, (1,)).item()
        x = torch.randn(N, C, 64, 64, dtype=torch.float32) * 10
        self._test_avg_pool_bf16_base(dim=2, input=x)

    def test_avg_pool3d_bf16(self):
        N = torch.randint(3, 10, (1,)).item()
        C = torch.randint(3, 10, (1,)).item()
        x = torch.randn(N, C, 64, 64, 64, dtype=torch.float32) * 10
        self._test_avg_pool_bf16_base(dim=3, input=x)

    def test_avg_pool2d_stride_none(self):
        N = torch.randint(3, 10, (1,)).item()
        C = torch.randint(3, 10, (1,)).item()
        x = torch.randn(N, C, 64, 64, dtype=torch.float32) * 10

        for count_include_pad in [True, False]:
            y1 = F.avg_pool2d(
                x,
                kernel_size=3,
                stride=None,
                padding=1,
                count_include_pad=count_include_pad)
            y2 = F.avg_pool2d(
                x.to_mkldnn(),
                kernel_size=3,
                stride=None,
                padding=1,
                count_include_pad=count_include_pad)

            self.assertEqual(y1, y2.to_dense())

    def test_adaptive_avg_pool2d(self):
        N = torch.randint(3, 10, (1,)).item()
        C = torch.randint(3, 10, (1,)).item()
        x = torch.randn(N, C, 224, 224, dtype=torch.float32) * 100

        adaptive_avg_pool2d = torch.nn.AdaptiveAvgPool2d(7)
        x1 = x.clone().requires_grad_()
        x2 = x.clone().to_mkldnn().requires_grad_()
        y1 = adaptive_avg_pool2d(x1)
        y2 = adaptive_avg_pool2d(x2).to_dense()

        loss1 = y1.sum()
        loss2 = y2.sum()
        loss1.backward()
        loss2.backward()

        self.assertEqual(y1, y2)
        self.assertEqual(x1.grad, x2.grad.to_dense())

    @unittest.skipIf(IS_WINDOWS, "Limit support for bf16 path")
    def test_adaptive_avg_pool2d_bf16(self):
        N = torch.randint(3, 10, (1,)).item()
        C = torch.randint(3, 10, (1,)).item()
        x = torch.randn(N, C, 224, 224, dtype=torch.float32) * 100

        x_bf16 = x.bfloat16()
        adaptive_avg_pool2d = torch.nn.AdaptiveAvgPool2d(7)

        if torch.ops.mkldnn._is_mkldnn_bf16_supported():
            y = adaptive_avg_pool2d(x.to_mkldnn()).to_dense()
            y_bf16 = adaptive_avg_pool2d(x.to_mkldnn()).to_dense(torch.float32)
            self.assertEqual(y, y_bf16, atol=1e-1, rtol=1e-3)
        else:
            msg = "mkldnn_adaptive_avg_pool2d: bf16 path needs the cpu support avx512bw, avx512vl and avx512dq"
            self.assertRaisesRegex(RuntimeError,
                                   msg,
                                   lambda: adaptive_avg_pool2d(x_bf16.to_mkldnn()))

    def _test_batch_norm_base(self, dim, channels, input):
        bn_module = {2 : torch.nn.BatchNorm2d, 3 : torch.nn.BatchNorm3d}
        bn = bn_module[dim](channels).float().train(False)
        mkldnn_bn = mkldnn_utils.to_mkldnn(copy.deepcopy(bn))
        self.assertEqual(
            bn(input),
            mkldnn_bn(input.to_mkldnn()).to_dense())

        self._test_serialization(mkldnn_bn, (input.to_mkldnn(),))
        self._test_tracing(mkldnn_bn, (input.to_mkldnn(),))

    def _test_batch_norm_train_base(self, dim, channels, input):
        # TODO: support 3d batchnorm training.
        bn_module = {2 : torch.nn.BatchNorm2d}
        # TODO: support none affine.
        options = itertools.product([True], [True, False])
        for affine, track_running_stats in options:
            bn = bn_module[dim](
                num_features=channels,
                affine=affine,
                track_running_stats=track_running_stats).float().train(True)
            mkldnn_bn = copy.deepcopy(bn)
            x1 = input.clone().requires_grad_()
            x2 = input.clone().to_mkldnn().requires_grad_()
            y1 = bn(x1)
            y2 = mkldnn_bn(x2).to_dense()
            loss1 = y1.sum()
            loss2 = y2.sum()
            loss1.backward()
            loss2.backward()
            self.assertEqual(y1, y2)
            self.assertEqual(x1.grad, x2.grad.to_dense())
            self.assertEqual(bn.weight.grad, mkldnn_bn.weight.grad, rtol=1e-3, atol=1e-3)
            if track_running_stats:
                self.assertEqual(bn.running_mean, mkldnn_bn.running_mean)
                self.assertEqual(bn.running_var, mkldnn_bn.running_var, rtol=1e-5, atol=1e-5)

    def test_batch_norm_2d(self):
        N = torch.randint(3, 10, (1,)).item()
        C = torch.randint(3, 100, (1,)).item()
        x = torch.randn(N, C, 35, 45, dtype=torch.float32) * 10
        self._test_batch_norm_base(dim=2, channels=C, input=x)
        self._test_batch_norm_train_base(dim=2, channels=C, input=x)

    def test_batch_norm_3d(self):
        N = torch.randint(3, 10, (1,)).item()
        C = torch.randint(3, 100, (1,)).item()
        x = torch.randn(N, C, 30, 30, 30, dtype=torch.float32) * 10
        self._test_batch_norm_base(dim=3, channels=C, input=x)

    @unittest.skipIf(IS_WINDOWS, "Limit support for bf16 path")
    def _test_batch_norm_bf16_base(self, dim, channels, input):
        bn_module = {2 : torch.nn.BatchNorm2d, 3 : torch.nn.BatchNorm3d}
        x_bf16 = input.bfloat16()
        # TODO: support training
        for train in [False]:
            bn = bn_module[dim](channels).float().train(train)
            mkldnn_bn = mkldnn_utils.to_mkldnn(copy.deepcopy(bn))  # noqa: F841
            if torch.ops.mkldnn._is_mkldnn_bf16_supported():
                y = bn(input.to_mkldnn().to_dense())
                y_bf16 = bn(input.to_mkldnn().to_dense(torch.float))
                self.assertEqual(y, y_bf16, atol=1e-1, rtol=1e-3)
            else:
                msg = "mkldnn_batch_norm: bf16 path needs the cpu support avx512bw, avx512vl and avx512dq"
                self.assertRaisesRegex(RuntimeError,
                                       msg,
                                       lambda: bn(x_bf16.to_mkldnn()))

    def test_batch_norm_2d_bf16(self):
        N = torch.randint(3, 10, (1,)).item()
        C = torch.randint(3, 100, (1,)).item()
        x = torch.randn(N, C, 35, 45, dtype=torch.float32) * 10
        self._test_batch_norm_bf16_base(dim=2, channels=C, input=x)

    def test_batch_norm_3d_bf16(self):
        N = torch.randint(3, 10, (1,)).item()
        C = torch.randint(3, 100, (1,)).item()
        x = torch.randn(N, C, 30, 30, 30, dtype=torch.float32) * 10
        self._test_batch_norm_bf16_base(dim=3, channels=C, input=x)

    def test_add(self):
        N = torch.randint(3, 10, (1,)).item()
        C = torch.randint(3, 100, (1,)).item()
        alpha = torch.randn(1, dtype=torch.float32).item()

        x = torch.randn(N, C, 35, 45, dtype=torch.float32) * 10
        y = torch.randn(N, C, 35, 45, dtype=torch.float32) * 10
        mx = x.to_mkldnn()
        my = y.to_mkldnn()

        # add
        self.assertEqual(
            x + y,
            (mx + my).to_dense())

        self.assertEqual(
            torch.add(x, y, alpha=alpha),
            torch.add(mx, my, alpha=alpha).to_dense())

        # add_
        x += y
        mx += my
        self.assertEqual(x, mx.to_dense())

        # add_out
        out = x.clone()
        mkldnn_out = out.to_mkldnn()
        torch.add(x, y, alpha=alpha, out=out)
        torch.add(mx, my, alpha=alpha, out=mkldnn_out)
        self.assertEqual(out, mkldnn_out.to_dense())

        # add_out inplace case: first input
        torch.add(x, y, alpha=alpha, out=x)
        torch.add(mx, my, alpha=alpha, out=mx)
        self.assertEqual(x, mx.to_dense())

        # add_out inplace case: second input
        torch.add(x, y, alpha=alpha, out=y)
        torch.add(mx, my, alpha=alpha, out=my)
        self.assertEqual(y, my.to_dense())

    def test_mul(self):
        N = torch.randint(3, 10, (1,)).item()
        C = torch.randint(3, 100, (1,)).item()
        value = torch.randn(1, dtype=torch.float32).item()

        x = torch.randn(N, C, 35, 45, dtype=torch.float32) * 10
        y = torch.randn(N, C, 35, 45, dtype=torch.float32) * 10
        mx = x.to_mkldnn()
        my = y.to_mkldnn()

        # mul
        self.assertEqual(
            x * y,
            (mx * my).to_dense())

        self.assertEqual(
            x * value,
            (mx * value).to_dense())

        self.assertEqual(
            torch.mul(x, y),
            torch.mul(mx, my).to_dense())

        self.assertEqual(
            torch.mul(x, value),
            torch.mul(mx, value).to_dense())

        # mul_
        x *= y
        mx *= my
        self.assertEqual(x, mx.to_dense())

        x *= value
        mx *= value
        self.assertEqual(x, mx.to_dense())

        # mul_out
        out = x.clone()
        mkldnn_out = out.to_mkldnn()
        torch.mul(x, y, out=out)
        torch.mul(mx, my, out=mkldnn_out)
        self.assertEqual(out, mkldnn_out.to_dense())

        out = x.clone()
        mkldnn_out = out.to_mkldnn()
        torch.mul(x, value, out=out)
        torch.mul(mx, value, out=mkldnn_out)
        self.assertEqual(out, mkldnn_out.to_dense())

    def test_0_dimension_tensor(self):
        x = torch.rand([20, 20, 1, 1], dtype=torch.float)
        y = torch.rand([20, 20, 0, 1], dtype=torch.float)

        # unary ops work without modification
        out_relu = torch.relu(y)
        out_relu_mkldnn = torch.relu(y.to_mkldnn()).to_dense()
        self.assertEqual(out_relu, out_relu_mkldnn)

        out_mul = x * y
        out_mul_mkldnn = (x.to_mkldnn() * y.to_mkldnn()).to_dense()
        self.assertEqual(out_mul, out_mul_mkldnn)

        out_add = x + y
        out_add_mkldnn = (x.to_mkldnn() + y.to_mkldnn()).to_dense()
        self.assertEqual(out_add, out_add_mkldnn)

        x.requires_grad_(True)
        y.requires_grad_(True)
        with self.assertRaisesRegex(RuntimeError, "0-dimension Tensor in training"):
            x.to_mkldnn() + y.to_mkldnn()

        with self.assertRaisesRegex(RuntimeError, "must match"):
            torch.rand([5]).to_mkldnn() + torch.rand([0]).to_mkldnn()

        C = 7
        m = torch.nn.Conv2d(C, C, 3)
        x = torch.randn(0, C, C, 8, dtype=torch.float)
        out_eager = m(x)
        out_mkldnn = mkldnn_utils.to_mkldnn(m)(x)
        self.assertEqual(out_eager, out_mkldnn)

    # https://github.com/pytorch/pytorch/issues/127111
    @xfailIfTorchDynamo
    def test_view(self):
        x = torch.randn(3, 4, 5, dtype=torch.float32).to_mkldnn()
        self.assertRaisesRegex(RuntimeError,
                               "Change to use reshape",
                               lambda: x.view(x.size(0), -1))

    def test_reshape(self):
        x = torch.randn(3, 4, 5, dtype=torch.float32) * 10
        size = (x.size(0), -1)

        self.assertEqual(
            x.reshape(size),
            x.to_mkldnn().reshape(size).to_dense(),
        )
        # test whether share same memory for plain format tensor
        y = x.to_mkldnn()
        z = y.reshape(size).add_(y.reshape(size))
        self.assertEqual(
            y.reshape(size).to_dense(),
            z.to_dense(),
        )

    def test_reshape_blocked_format(self):
        # construct an mkldnn blocked tensor with mkldnn conv2d
        C = 7
        m = mkldnn_utils.to_mkldnn(torch.nn.Conv2d(C, C, 3))
        x = torch.randn(1, C, 8, 8).to_mkldnn()

        # mkldnn tensor w/ blocked format
        y_block = m(x)
        # aten tensor w/ plain format
        y_plain = y_block.to_dense()

        y_block_reshape = y_block.reshape(C, -1)
        y_plain_reshape = y_plain.reshape(C, -1)

        self.assertEqual(y_plain_reshape, y_block_reshape.to_dense())

    def test_reshape_backward(self):
        x = torch.randn(3, 4, 5, dtype=torch.float32) * 10
        size = (x.size(0), -1)

        x1 = x.clone().requires_grad_()
        x2 = x.clone().to_mkldnn().requires_grad_()
        in_features = 20
        out_features = torch.randint(3, 100, (1,)).item()
        linear = torch.nn.Linear(in_features, out_features).float()

        y1 = linear(x1.reshape(size)).sum()
        y2 = linear(x2.reshape(size).to_dense()).sum()
        y1.backward()
        y2.backward()
        self.assertEqual(x1.grad, x2.grad.to_dense())

    def test_clone(self):
        x = torch.randn(4, 5, dtype=torch.float32) * 10
        self.assertEqual(
            x.clone(),
            x.to_mkldnn().clone().to_dense(),
        )
        # test whether share same memory
        y = x.to_mkldnn()
        z = y.clone().add_(y)
        self.assertNotEqual(
            y.to_dense(),
            z.to_dense(),
        )

    def test_transpose(self):
        x = torch.randn(3, 4, 5, dtype=torch.float32) * 10
        for dim1 in range(x.ndim):
            for dim2 in range(x.ndim):
                self.assertEqual(
                    x.transpose(dim1, dim2),
                    x.to_mkldnn().transpose(dim1, dim2).to_dense(),
                )

    def test_transpose_invalid_dime(self):
        x = torch.randn(3, 4, 5, dtype=torch.float32).to_mkldnn()
        with self.assertRaisesRegex(IndexError, "Dimension out of range"):
            torch._mkldnn_transpose(x, 0, 12)

    def test_linear_non_contiguous_weight(self):
        in_features = torch.randint(3, 10, (1,)).item()
        out_features = torch.randint(3, 100, (1,)).item()
        x = torch.randn(3, in_features, dtype=torch.float32) * 10
        w = torch.randn(in_features, out_features, dtype=torch.float32)
        for bias in [True, False]:
            x1 = x.clone().requires_grad_()
            x2 = x.clone().to_mkldnn().requires_grad_()
            linear = torch.nn.Linear(in_features, out_features).float()
            linear.weight = torch.nn.Parameter(w.t())
            mkldnn_linear = copy.deepcopy(linear)
            y1 = linear(x1).sum()
            y2 = mkldnn_linear(x2).to_dense().sum()
            y1.backward()
            y2.backward()
            self.assertEqual(x1.grad, x2.grad.to_dense())
            self.assertEqual(linear.weight.grad, mkldnn_linear.weight.grad)
            if bias:
                self.assertEqual(linear.bias.grad, mkldnn_linear.bias.grad)

    def test_linear(self):
        in_features = torch.randint(3, 10, (1,)).item()
        out_features = torch.randint(3, 100, (1,)).item()
        x = torch.randn(3, in_features, dtype=torch.float32) * 10

        for bias in [True, False]:
            linear = torch.nn.Linear(in_features, out_features, bias=bias).float()
            mkldnn_linear = mkldnn_utils.to_mkldnn(copy.deepcopy(linear))
            self.assertEqual(
                linear(x),
                mkldnn_linear(x.to_mkldnn()).to_dense())

            self._test_serialization(mkldnn_linear, (x.to_mkldnn(),))
            self._test_tracing(mkldnn_linear, (x.to_mkldnn(),))

    def test_linear_backward(self):
        in_features = torch.randint(3, 10, (1,)).item()
        out_features = torch.randint(3, 100, (1,)).item()
        x = torch.randn(3, in_features, dtype=torch.float32) * 10
        for bias in [True, False]:
            x1 = x.clone().requires_grad_()
            x2 = x.clone().to_mkldnn().requires_grad_()
            linear = torch.nn.Linear(in_features, out_features).float()
            mkldnn_linear = copy.deepcopy(linear)
            y1 = linear(x1).sum()
            y2 = mkldnn_linear(x2).to_dense().sum()
            y1.backward()
            y2.backward()
            self.assertEqual(x1.grad, x2.grad.to_dense())
            self.assertEqual(linear.weight.grad, mkldnn_linear.weight.grad)
            if bias:
                self.assertEqual(linear.bias.grad, mkldnn_linear.bias.grad)

    @dtypes(torch.float16, torch.bfloat16)
    def test_linear_lowp(self, dtype):
        in_features = torch.randint(3, 10, (1,)).item()
        out_features = torch.randint(3, 100, (1,)).item()
        x = torch.randn(3, in_features, dtype=torch.float32) * 10
        x_lowp = x.to(dtype=dtype)

        for bias in [True, False]:
            linear = torch.nn.Linear(in_features, out_features, bias=bias).float()
            mkldnn_linear = mkldnn_utils.to_mkldnn(copy.deepcopy(linear))
            mkldnn_linear_lowp = mkldnn_utils.to_mkldnn(
                copy.deepcopy(linear), dtype
            )
            lowp_support = {
                torch.bfloat16: torch.ops.mkldnn._is_mkldnn_bf16_supported,
                torch.half: torch.ops.mkldnn._is_mkldnn_fp16_supported,
            }
            if lowp_support[dtype]():
                y = mkldnn_linear(x.to_mkldnn()).to_dense()
                y_lowp = mkldnn_linear_lowp(x_lowp.to_mkldnn()).to_dense(
                    torch.float32
                )
                if dtype == torch.bfloat16:
                    self.assertEqual(y, y_lowp, atol=1e-1, rtol=1e-3)
                else:
                    self.assertEqual(y, y_lowp, atol=5e-3, rtol=1e-3)
            else:
                msg = {
                    torch.bfloat16: r"bf16 path needs the cpu support avx_ne_convert or avx512bw, avx512vl and avx512dq",
                    torch.half: r"fp16 path needs the cpu support avx_ne_convert or avx512_fp16",
                }
                self.assertRaisesRegex(
                    RuntimeError,
                    msg[dtype],
                    lambda: mkldnn_linear_lowp(x_lowp.to_mkldnn()),
                )

    def test_softmax(self):
        x = torch.randn(3, 4, 5, dtype=torch.float32) * 10
        for dim in range(x.ndim):
            softmax = torch.nn.Softmax(dim=dim)
            self.assertEqual(
                softmax(x),
                softmax(x.to_mkldnn()).to_dense())

    def test_sigmoid(self):
        x = torch.randn(4, 5, dtype=torch.float32) * 10
        mkldnn_x = x.to_mkldnn()
        self.assertEqual(
            torch.sigmoid(x),
            torch.sigmoid(mkldnn_x).to_dense(),
        )
        # inplace
        torch.sigmoid_(x)
        torch.sigmoid_(mkldnn_x)
        self.assertEqual(x, mkldnn_x.to_dense())

    def test_tanh(self):
        x = torch.randn(4, 5, dtype=torch.float32) * 10
        mkldnn_x = x.to_mkldnn()
        self.assertEqual(
            torch.tanh(x),
            torch.tanh(mkldnn_x).to_dense(),
        )
        # inplace
        torch.tanh_(x)
        torch.tanh_(mkldnn_x)
        self.assertEqual(x, mkldnn_x.to_dense())

    def _test_serialization(self, module, inputs):
        with TemporaryFileName() as fname:
            torch.jit.save(module, fname)
            loaded = torch.jit.load(fname)
            self.assertEqual(
                module(*inputs).to_dense(),
                loaded(*inputs).to_dense())

    def _test_tracing(self, module, inputs):
        traced = torch.jit.trace(module, inputs)
        self.assertEqual(
            module(*inputs).to_dense(),
            traced(*inputs).to_dense())

    def test_set_data_tensorimpl_type(self):
        # Dense tensor has impl of type `TensorImpl`, while MKL-DNN tensor has impl
        # of type `OpaqueTensorImpl<IDeepTensorWrapperPtr>`.
        x = torch.randn((1, 2), dtype=torch.float, device=torch.device('cpu'))
        x_mkldnn = x.to_mkldnn()
        with self.assertRaisesRegex(RuntimeError, 'incompatible tensor type'):
            x.data = x_mkldnn

    def test_empty(self):
        x1 = torch.empty(4, 5, 2, 3, dtype=torch.float32)
        x2 = torch.empty(4, 5, 2, 3, dtype=torch.float32, layout=torch._mkldnn)
        self.assertEqual(x1.size(), x2.to_dense().size())
        self.assertEqual(x1.dtype, x2.to_dense().dtype)

    def test_zero_(self):
        x1 = torch.randn(4, 5, dtype=torch.float32) * 10
        x2 = x1.clone().to_mkldnn()
        self.assertEqual(
            x1.zero_(),
            x2.zero_().to_dense(),
        )

    def test_is_mkldnn(self):
        x = torch.randn(1, dtype=torch.float32)
        self.assertFalse(x.is_mkldnn)
        self.assertTrue(x.to_mkldnn().is_mkldnn)

    # legacy constructor/new doesn't support mkldnn tensors
    @skipIfTorchDynamo("https://github.com/pytorch/torchdynamo/issues/1992")
    def test_legacy_new_failure(self):
        x = torch.randn(1, dtype=torch.float32)
        x_mkldnn = x.to_mkldnn()
        self.assertRaises(RuntimeError, lambda: x_mkldnn.new(device='cpu'))
        self.assertRaises(RuntimeError, lambda: x_mkldnn.new(x.storage()))
        self.assertRaises(RuntimeError, lambda: x_mkldnn.new(x))
        self.assertRaises(RuntimeError, lambda: x_mkldnn.new(torch.Size([2, 3])))
        self.assertRaises(RuntimeError, lambda: x_mkldnn.new([6]))

    def test_is_mkldnn_jit(self):
        class EnsureMkldnn(torch.jit.ScriptModule):
            @torch.jit.script_method
            def forward(self, x):
                if not x.is_mkldnn:
                    x = x.to_mkldnn()
                return x

        m = EnsureMkldnn()
        x = torch.randn(1, dtype=torch.float32)
        self.assertTrue(m(x).is_mkldnn)
        self.assertTrue(m(x.to_mkldnn()).is_mkldnn)

    def _test_imagenet_model(self, model):
        model = model.train(False).float()
        mkldnn_model = mkldnn_utils.to_mkldnn(copy.deepcopy(model))
        x = torch.randn(1, 3, 224, 224, dtype=torch.float32)
        with torch.no_grad():
            self.assertEqual(
                model(x),
                mkldnn_model(x.to_mkldnn()).to_dense(),
            )

    @skipIfNoTorchVision
    def test_resnet18(self):
        model = torchvision.models.resnet.resnet18(weights=None)
        self._test_imagenet_model(model)

    @skipIfNoTorchVision
    def test_resnext50_32x4d(self):
        model = torchvision.models.resnet.resnext50_32x4d(weights=None)
        self._test_imagenet_model(model)

    def _lstm_params_list(self):
        params_dict = {
            "input_size": [1, 5],
            "hidden_size": [5, 16],
            "num_layers": [1, 3],
            "bidirectional": [False, True],
            "bias": [False, True],
            "batch_first": [False, True],
            "dropout": [0, 0.4, 0.7, 1],
            "batch_size": [1, 2],
            "seq_len": [1, 3],
            "training": [False, True]
        }

        params_list = list(params_dict.values())
        return params_list

    def _cast_dtype(self, input, dtype):
        if dtype == torch.bfloat16:
            input = input.to(torch.bfloat16)
        elif dtype == torch.half:
            input = input.to(torch.half)
        return input

    def test_lstm(self):
        seed = 2023
        torch.manual_seed(seed)

        params_list = self._lstm_params_list()
        for dtype in types:
            bf16 = dtype == torch.bfloat16
            fp16 = dtype == torch.half
            rtol = 1.3e-6
            atol = 1e-5

            if bf16:
                rtol = 0.02
                atol = 0.02
            if fp16:
                rtol = 1e-3
                atol = 1e-3
            for input_size, hidden_size, num_layers, bidirectional, bias, batch_first, dropout, batch_size, seq_len, training \
                    in itertools.product(*params_list):
                num_directions = 2 if bidirectional else 1
                if batch_first:
                    input = torch.randn(batch_size, seq_len, input_size, dtype=torch.float32)
                else:
                    input = torch.randn(seq_len, batch_size, input_size, dtype=torch.float32)
                h = torch.randn(num_layers * num_directions, batch_size, hidden_size, dtype=torch.float32)
                c = torch.randn(num_layers * num_directions, batch_size, hidden_size, dtype=torch.float32)
                if fp16:
                    # TODO add traing support when oneDNN support lstm FP16 training
                    training = False
                model = torch.nn.LSTM(input_size, hidden_size, num_layers, bidirectional=bidirectional,
                                      bias=bias, dropout=dropout, batch_first=batch_first).float()
                model.train() if training else model.eval()
                input1 = input.clone().requires_grad_(training)
                input2 = input.clone().requires_grad_(training)

                h1 = h.clone().requires_grad_(training)
                h2 = h.clone().requires_grad_(training)
                c1 = c.clone().requires_grad_(training)
                c2 = c.clone().requires_grad_(training)

                model1 = copy.deepcopy(model)
                model2 = copy.deepcopy(model)
                with torch.no_grad() if not training else nullcontext():
                    with torch.backends.mkldnn.flags(enabled=False):
                        torch.manual_seed(seed)
                        output1, (hn1, cn1) = self._cast_dtype(model1, dtype)(
                            self._cast_dtype(input1, dtype),
                            (
                                self._cast_dtype(h1, dtype),
                                self._cast_dtype(c1, dtype),
                            ),
                        )

                    torch.manual_seed(seed)
                    output2, (hn2, cn2) = self._cast_dtype(model2, dtype)(
                        self._cast_dtype(input2, dtype),
                        (
                            self._cast_dtype(h2, dtype),
                            self._cast_dtype(c2, dtype),
                        ),
                    )
                    self.assertEqual(output1, output2, rtol=rtol, atol=atol)
                    self.assertEqual(hn1, hn2, rtol=rtol, atol=atol)
                    self.assertEqual(cn1, cn2, rtol=rtol, atol=atol)

                    if training:
                        with torch.backends.mkldnn.flags(enabled=False):
                            torch.manual_seed(seed)
                            output1.sum().backward(retain_graph=True)

                        torch.manual_seed(seed)
                        output2.sum().backward(retain_graph=True)

                        self.assertEqual(input1.grad, input2.grad, rtol=rtol, atol=atol)
                        for name, para in model1.named_parameters():
                            self.assertEqual(para, getattr(model2, name))
                            self.assertEqual(
                                para.grad,
                                getattr(model2, name).grad,
                                rtol=rtol,
                                atol=atol,
                            )

                        with torch.backends.mkldnn.flags(enabled=False):
                            torch.manual_seed(seed)
                            hn1.sum().backward(retain_graph=True)
                        torch.manual_seed(seed)
                        hn2.sum().backward(retain_graph=True)
                        self.assertEqual(h1.grad, h2.grad, rtol=rtol, atol=atol)

                        with torch.backends.mkldnn.flags(enabled=False):
                            torch.manual_seed(seed)
                            cn1.sum().backward(retain_graph=True)
                        torch.manual_seed(seed)
                        cn2.sum().backward(retain_graph=True)
                        self.assertEqual(c1.grad, c2.grad, rtol=rtol, atol=atol)

    @dtypes(torch.float16, torch.bfloat16)
    def test_matmul_lower_precision(self, dtype):
        support_check = {
            torch.bfloat16: torch.ops.mkldnn._is_mkldnn_bf16_supported,
            torch.float16: torch.ops.mkldnn._is_mkldnn_fp16_supported,
        }

        def common(self, shape1, shape2, op, dtype):
            a = torch.randn(shape1, dtype=dtype)
            a_ref = a.float()
            b = torch.randn(shape2, dtype=dtype)
            b_ref = b.float()

            y = op(a, b)
            y_ref = op(a_ref, b_ref)
            self.assertEqual(y, y_ref, exact_dtype=False)

        if support_check[dtype]():
            a1 = torch.randn([64, 1, 33], dtype=dtype)
            # a2 is contiguous tensor but it's strides
            # is not default contiguous strides.
            a2 = torch.as_strided(a1.clone(), [64, 1, 33], [33, 3, 1])
            self.assertTrue(a2.is_contiguous())
            b = torch.randn(64, 33, 256).to(dtype=dtype)
            y1 = torch.ops.aten.bmm(a1, b)
            y2 = torch.bmm(a2, b)
            self.assertEqual(y1, y2)

            for shape1, shape2, op in [
                ((33, 77), (77, 22), torch.matmul),
                ((128, 256), (256, 10), torch.matmul),
                ((7, 300), (300, 3), torch.matmul),
                ((1, 100), (100, 60), torch.matmul),
                ((100, 1), (1, 100), torch.matmul),
                ((20, 54, 78), (20, 78, 10), torch.bmm),
                ((1, 300, 1), (1, 1, 300), torch.bmm),
            ]:
                common(self, shape1, shape2, op, dtype)


instantiate_device_type_tests(TestMkldnn, globals(), only_for=('cpu',))

if __name__ == '__main__':
    run_tests()