enable DNNL Python OPs(adaptive_avg_pool2d, max_pool2d, max_pool3d) to fallback to CPU.

Author: jiayisunx

intel_pytorch_extension_py/ops/linear.py

@@ -24,7 +24,7 @@ def backward(ctx, grad_output):
         return (grad_input, grad_weight, grad_bias)
 
 def linear(input, weight, bias=None):
-    if input.device.type == 'dpcpp':
+    if input.device.type == 'dpcpp' and core.get_auto_dnnl():
         return LinearFunction.apply(input, weight, bias)
     return F_linear(input, weight, bias)
 

intel_pytorch_extension_py/ops/pooling.py

@@ -3,15 +3,16 @@
 import torch.nn.functional as F
 import _torch_ipex as core
 
-F_adaptive_avg_pool2d = F.adaptive_avg_pool2d
+torch_adaptive_avg_pool2d = torch._C._nn.adaptive_avg_pool2d
 torch_max_pool2d = torch.max_pool2d
 torch_max_pool3d = torch.max_pool3d
 
 class AdaptiveAvgPool2dFunction(Function):
     @staticmethod
     def forward(ctx, input, output_size):
-        _output_size = _list_with_default(output_size, input.size())
-        output = core.adaptive_avg_pool2d(input, _output_size)
+        if type(output_size) is int:
+            output_size = (output_size,)
+        output = core.adaptive_avg_pool2d(input, output_size)
         ctx.save_for_backward(input)
         return output
 
@@ -25,7 +26,15 @@ def backward(ctx, grad_output):
 class MaxPoolingFunction(Function):
     @staticmethod
     def forward(ctx, input, kernel_size, stride, padding, dilation, ceil_mode):
-        output = core.max_pooling(input, (kernel_size,), (stride,), (padding,), (dilation,), ceil_mode)
+        if type(kernel_size) is int:
+            kernel_size = (kernel_size,)
+        if type(stride) is int:
+            stride = (stride,)
+        if type(padding) is int:
+            padding = (padding,)
+        if type(dilation) is int:
+            dilation = (dilation,)
+        output = core.max_pooling(input, kernel_size, stride, padding, dilation, ceil_mode)
         ctx.save_for_backward(output, input)
         ctx.kernel_size = kernel_size
         ctx.stride = stride
@@ -38,31 +47,33 @@ def forward(ctx, input, kernel_size, stride, padding, dilation, ceil_mode):
     def backward(ctx, grad_output):
         output, input= ctx.saved_tensors
         grad_output = grad_output.contiguous()
-        grad_input = core.max_pooling_backward(grad_output, output, input, (ctx.kernel_size,), (ctx.stride,), (ctx.padding,), (ctx.dilation,), ctx.ceil_mode)
+        grad_input = core.max_pooling_backward(grad_output, output, input, ctx.kernel_size, ctx.stride, ctx.padding, ctx.dilation, ctx.ceil_mode)
         return (grad_input, None, None, None, None, None)
 
-def _list_with_default(out_size, defaults):
-    if isinstance(out_size, int):
-        return (out_size,)
-    if len(defaults) <= len(out_size):
-        raise ValueError('Input dimension should be at least {}'.format(len(out_size) + 1))
-    return [v if v is not None else d for v, d in zip(out_size, defaults[-len(out_size):])]
-
 def adaptive_avg_pool2d(input, output_size):
-    if input.device.type == 'dpcpp':
-        return AdaptiveAvgPool2dFunction.apply(input, output_size)
-    return F_adaptive_avg_pool2d(input, output_size)
+    try:
+        if input.device.type == 'dpcpp' and core.get_auto_dnnl():
+            return AdaptiveAvgPool2dFunction.apply(input, output_size)
+    except RuntimeError:
+        return torch_adaptive_avg_pool2d(input, output_size)
+    return torch_adaptive_avg_pool2d(input, output_size)
 
 def max_pool2d(input, kernel_size, stride, padding, dilation, ceil_mode):
-    if input.device.type == 'dpcpp':
-        return MaxPoolingFunction.apply(input, kernel_size, stride, padding, dilation, ceil_mode)
+    try:
+        if input.device.type == 'dpcpp' and core.get_auto_dnnl():
+            return MaxPoolingFunction.apply(input, kernel_size, stride, padding, dilation, ceil_mode)
+    except RuntimeError:
+        return torch_max_pool2d(input, kernel_size, stride, padding, dilation, ceil_mode)
     return torch_max_pool2d(input, kernel_size, stride, padding, dilation, ceil_mode)
 
 def max_pool3d(input, kernel_size, stride, padding, dilation, ceil_mode):
-    if input.device.type == 'dpcpp':
-        return MaxPoolingFunction.apply(input, kernel_size, stride, padding, dilation, ceil_mode)
+    try:
+        if input.device.type == 'dpcpp' and core.get_auto_dnnl():
+            return MaxPoolingFunction.apply(input, kernel_size, stride, padding, dilation, ceil_mode)
+    except RuntimeError:
+        return torch_max_pool3d(input, kernel_size, stride, padding, dilation, ceil_mode)
     return torch_max_pool3d(input, kernel_size, stride, padding, dilation, ceil_mode)
 
-F.adaptive_avg_pool2d = adaptive_avg_pool2d
+torch._C._nn.adaptive_avg_pool2d = adaptive_avg_pool2d
 torch.max_pool2d = max_pool2d
 torch.max_pool3d = max_pool3d

intel_pytorch_extension_py/ops/reshape.py

@@ -11,7 +11,7 @@ def forward(ctx, input, size):
         return output
 
 def reshape(input, size):
-    if input.device.type == 'dpcpp':
+    if input.device.type == 'dpcpp' and core.get_auto_dnnl():
         return ReshapeFunction.apply(input, size)
     return torch_reshape(input, size)