Commit59ca444

authored

Merge pull request#23560 from WanliZhong:eltwise_cuda_bug

DNN/CUDA: Solve the bug of same shape broadcast with CUDA

2 parents04d71da +46991bc commit59ca444Copy full SHA for 59ca444

File tree

-4

lines changed

-4

lines changed

Lines changed: 4 additions & 1 deletion

Original file line number	Diff line number	Diff line change
`@@ -150,7 +150,7 @@ void eltwise_op(const Stream& stream, TensorSpan<T> output, TensorView<T> x, Ten`
`150`	`150`	`*/`
`151`	`151`	`for (int r =0; r < output.rank(); r++)`
`152`	`152`	`{`
`153`		`-while (x.get_axis_size(r) ==1 && y.get_axis_size(r) ==1) {`
	`153`	`+while (x.rank() > r && y.rank() > r && x.get_axis_size(r) ==1 && y.get_axis_size(r) ==1) {`
`154`	`154`	`CV_Assert(output.get_axis_size(r) ==1);`
`155`	`155`
`156`	`156`	`x.squeeze(r);`
`@@ -183,6 +183,9 @@ void eltwise_op(const Stream& stream, TensorSpan<T> output, TensorView<T> x, Ten`
`183`	`183`	`auto new_size = inShape1[i] * inShape1[j];`
`184`	`184`	`inShape1[i] = new_size;`
`185`	`185`	`inShape2[i] = new_size;`
	`186`	`+// outShape should be changed after merged`
	`187`	`+auto output_new_size = outShape[i] * outShape[j];`
	`188`	`+ outShape[i] = output_new_size;`
`186`	`189`
`187`	`190`	/* delete axis `j`*/
`188`	`191`	`inShape1.erase(std::begin(inShape1) + j);`

Lines changed: 0 additions & 2 deletions

Original file line number	Diff line number	Diff line change
`@@ -51,8 +51,6 @@ class NaryEltwiseLayerImpl CV_FINAL : public NaryEltwiseLayer`
`51`	`51`	`WHERE,`
`52`	`52`	`} op;`
`53`	`53`
`54`		`-// If the eltwise implementation is modified, you need to force enable the 'Layer_Test_Eltwise_bcast'`
`55`		`-// test in the 'test_layers.cpp' file to make sure it all passes`
`56`	`54`	`NaryEltwiseLayerImpl(const LayerParams& params)`
`57`	`55`	`{`
`58`	`56`	`setParamsFrom(params);`

Lines changed: 4 additions & 1 deletion

Original file line number	Diff line number	Diff line change
`@@ -1949,6 +1949,9 @@ struct Layer_Test_Eltwise_bcast : testing::TestWithParam<tuple<string, int, tupl`
`1949`	`1949`	`int backend = get<0>(backend_target);`
`1950`	`1950`	`int target = get<1>(backend_target);`
`1951`	`1951`
	`1952`	`+if (backend == DNN_BACKEND_CUDA && dim >4)`
	`1953`	`+applyTestTag(CV_TEST_TAG_LONG);`
	`1954`	`+`
`1952`	`1955`	`vector<vector<int>> dim_shape_list;`
`1953`	`1956`	`get_all_arr(dim_shape_list, dim);`
`1954`	`1957`	`replace(dim_shape_list,1,3);`
`@@ -2028,7 +2031,7 @@ struct Layer_Test_Eltwise_bcast : testing::TestWithParam<tuple<string, int, tupl`
`2028`	`2031`	`}`
`2029`	`2032`	`};`
`2030`	`2033`
`2031`		`-TEST_P(Layer_Test_Eltwise_bcast,DISABLED_brute_force)`
	`2034`	`+TEST_P(Layer_Test_Eltwise_bcast,brute_force)`
`2032`	`2035`	`{`
`2033`	`2036`	`test_bcast();`
`2034`	`2037`	`}`

Comments

(0)