pytorch
diff --git a/‎aten/src/ATen/native/AdaptiveAveragePooling.cpp‎
Lines changed: 2 additions & 1 deletion b/‎aten/src/ATen/native/AdaptiveAveragePooling.cpp‎
Lines changed: 2 additions & 1 deletion
@@ -326,7 +326,8 @@ namespace {
       return at::mkldnn_adaptive_avg_pool2d(input, output_size);
     }
 
-    if (!input.is_quantized() && output_size[0] == 1 && output_size[1] == 1) {
+    // TODO: fastpath for Channels_last should be explored later;
+    if (input.suggest_memory_format() == at::MemoryFormat::Contiguous && !input.is_quantized() && output_size[0] == 1 && output_size[1] == 1) {
       // in this case, adaptive pooling is just computing mean over hw
       // dimensions, which can be done more efficiently
       int64_t mean_size = input.size(-1) * input.size(-2);
Original file line number	Diff line number	Diff line change
`@@ -326,7 +326,8 @@ namespace {`
`326`	`326`	`return at::mkldnn_adaptive_avg_pool2d(input, output_size);`
`327`	`327`	`}`
`328`	`328`
`329`		`- if (!input.is_quantized() && output_size[0] == 1 && output_size[1] == 1) {`
	`329`	`+ // TODO: fastpath for Channels_last should be explored later;`
	`330`	`+ if (input.suggest_memory_format() == at::MemoryFormat::Contiguous && !input.is_quantized() && output_size[0] == 1 && output_size[1] == 1) {`
`330`	`331`	`// in this case, adaptive pooling is just computing mean over hw`
`331`	`332`	`// dimensions, which can be done more efficiently`
`332`	`333`	`int64_t mean_size = input.size(-1) * input.size(-2);`