From 5b19383769ccf78f886f6df0db50e670f1bef0bc Mon Sep 17 00:00:00 2001
From: Xiaohan Ding <55726946+DingXiaoH@users.noreply.github.com>
Date: Sat, 30 Apr 2022 01:06:21 +0800
Subject: [PATCH] Update depthwise_conv2d_implicit_gemm.py

Sometimes when the tensor format changes after this conv (e.g., NCHW -> NHWC for layer normalization), calling backward will raise an "input must be contiguous" error. Making the grad contiguous in advance simply solves it.
---
 .../depthwise_conv2d_implicit_gemm.py                           | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/examples/19_large_depthwise_conv2d_torch_extension/depthwise_conv2d_implicit_gemm.py b/examples/19_large_depthwise_conv2d_torch_extension/depthwise_conv2d_implicit_gemm.py
index b4a1aa1c8a..f8c98b3344 100755
--- a/examples/19_large_depthwise_conv2d_torch_extension/depthwise_conv2d_implicit_gemm.py
+++ b/examples/19_large_depthwise_conv2d_torch_extension/depthwise_conv2d_implicit_gemm.py
@@ -21,6 +21,7 @@ def forward(ctx, x, w):
     @staticmethod
     @torch.cuda.amp.custom_bwd
     def backward(ctx, grad):
+        grad = grad.contiguous()
         x, w = ctx.saved_tensors
         dx = _extension.backward_data_fp32(grad, w)
         dw = _extension.backward_filter_fp32(grad, x, w)
@@ -37,6 +38,7 @@ def forward(ctx, x, w):
     @staticmethod
     @torch.cuda.amp.custom_bwd
     def backward(ctx, grad):
+        grad = grad.contiguous()
         x, w = ctx.saved_tensors
         dx = _extension.backward_data_fp16(grad, w)
         dw = _extension.backward_filter_fp16(grad, x, w)