Searched full:depthwise (Results 1 – 12 of 12) sorted by relevance
9 // Decompose TOSA Depthwise operation to a series of TOSA Ops specifically10 // (1) Convert a 1x1 Depthwise to Reshape -> Mul -> Reshape -> Add
22 // Masked vectorisation of 1D depthwise CW convs is not yet supported
4 // "depth" dimension in depthwise convs) followed by vectorization in the
1186 which is a special case for most depthwise convolutions.1207 which is a special case for most depthwise convolutions.1248 which is a special case for most depthwise convolutions.1269 which is a special case for most depthwise convolutions.1367 which is a special case for most depthwise convolutions.1402 which is a special case for most depthwise convolutions.
1821 // Support dynamic shapes in 1D depthwise convolution, but only in the in vectorizeLinalgOpPrecondition() 3412 return rewriter.notifyMatchFailure(op, "unvectorizable depthwise conv"); in depthwiseConv1dSliceAsMulAcc() 3746 op, "failed to match depthwise::Nwc conv 3-par 1-red");3754 return rewriter.notifyMatchFailure(op, "not a depthwise::Nwc layout");3849 // Only depthwise 1D NWC convs are left - these can be vectorized using masks3859 "Not a 1D depthwise conv!");
4200 which is a special case for most depthwise convolutions.4283 which is a special case for most depthwise convolutions.4451 which is a special case for most depthwise convolutions.4544 which is a special case for most depthwise convolutions.4993 which is a special case for most depthwise convolutions.5092 which is a special case for most depthwise convolutions.
1425 /// Rewrites 2-D depthwise convolution ops with size-1 (w, kw) or (h, kh)1426 /// dimensions into 1-D depthwise convolution ops.
223 // Creates a map to collapse the last dimension of the Depthwise convolution op in createDepthwiseConvCollapseMap()
960 // Depthwise loop. Already seen. in verifyFillInterface()
157 let summary = "Depthwise 2D Convolution operator";
235 (depthwise) convolutions, into combinations of lower-dimensional equivalents
3468 /// depthwise convolutions. This should lead to bette vectorization for in applyToOne()