PyTorch operator support for the Ethos-U85 backend

This page lists Ethos-U85-supported PyTorch APIs and the dtype and quantization modes covered by the Ethos-U85 backend test pipeline.

8x8 means 8-bit activations and 8-bit weights. 16x8 means 16-bit activations and 8-bit weights. 8x4 means 8-bit activations and 4-bit weights.

Total supported PyTorch APIs: 124.

PyTorch API	Support profile	DType	Quantization mode
torch.abs	INT	INT8	8x8
torch.acos	INT	INT8	8x8
torch.acosh	INT	INT8	8x8
torch.add / +	INT	INT8, INT16	8x8, 16x8
torch.addmm	INT	INT16	16x8
torch.alias_copy	INT	INT8	8x8
torch.amax	INT	INT8, INT16	8x8, 16x8
torch.amin	INT	INT8, INT16	8x8, 16x8
torch.any	INT	INT8	8x8
torch.arange	INT	INT8	8x8
torch.as_strided_copy	INT	INT8	8x8
torch.asin	INT	INT8	8x8
torch.asinh	INT	INT8	8x8
torch.atan	INT	INT8	8x8
torch.atanh	INT	INT8	8x8
torch.bitwise_and / &	INT	INT8	8x8
torch.bitwise_left_shift / <<	INT	INT8	8x8
torch.bitwise_not / ~	INT	INT8	8x8
torch.bitwise_or / |	INT	INT8	8x8
torch.bitwise_right_shift / >>	INT	INT8	8x8
torch.bitwise_xor / ^	INT	INT8	8x8
torch.bmm	INT	INT8, INT16	8x8, 16x8
torch.cat	INT	INT8, INT16	8x8, 16x8
torch.ceil	INT	INT8	8x8
torch.clamp	INT	INT8, INT16	8x8, 16x8
torch.clone / torch.Tensor.clone	INT	INT8	8x8
torch.cond	INT	INT8	8x8
torch.conv1d	INT	INT8, INT16	8x8, 16x8
torch.conv2d	INT	INT8, INT16, INT4	8x8, 8x4, 16x8
torch.conv3d	INT	INT8, INT4	8x8, 8x4
torch.conv_transpose2d	INT	INT8	8x8
torch.cos	INT	INT8	8x8
torch.cosh	INT	INT8	8x8
torch.cumsum	INT	INT8	8x8
torch.div	INT	INT8	8x8
torch.div / /	INT	INT8	8x8
torch.eq / ==	INT	INT8, INT16	8x8, 16x8
torch.erf	INT	INT8	8x8
torch.erfinv	INT	INT8	8x8
torch.exp	INT	INT8, INT16	8x8, 16x8
torch.expm1	INT	INT8	8x8
torch.eye	INT	INT8	8x8
torch.fill_	INT	INT8	8x8
torch.flip	INT	INT8	8x8
torch.floor	INT	INT8	8x8
torch.full	INT	INT8	8x8
torch.full_like	INT	INT8	8x8
torch.gather	INT	INT8	8x8
torch.ge / >=	INT	INT8, INT16	8x8, 16x8
torch.gt / >	INT	INT8, INT16	8x8, 16x8
torch.index_put_	INT	INT8	8x8
torch.index_select	INT	INT8	8x8
torch.le / <=	INT	INT8, INT16	8x8, 16x8
torch.linspace	INT	INT8	8x8
torch.log	INT	INT8	8x8
torch.log10	INT	INT8	8x8
torch.log1p	INT	INT8	8x8
torch.log_softmax	INT	INT8	8x8
torch.logical_and	INT	INT8	8x8
torch.logical_not	INT	INT8	8x8
torch.logical_or	INT	INT8	8x8
torch.logical_xor	INT	INT8	8x8
torch.lt / <	INT	INT8, INT16	8x8, 16x8
torch.masked_fill / torch.Tensor.masked_fill	INT	INT8	8x8
torch.max_pool2d	INT	INT8, INT16	8x8, 16x8
torch.maximum	INT	INT8, INT16	8x8, 16x8
torch.mean	INT	INT8	8x8
torch.minimum	INT	INT8, INT16	8x8, 16x8
torch.mm	INT	INT8	8x8
torch.mul / *	INT	INT8, INT16	8x8, 16x8
torch.neg / unary -	INT	INT8, INT16	8x8, 16x8
torch.nn.AvgPool2d / torch.nn.functional.avg_pool2d	INT	INT8, INT16	8x8, 16x8
torch.nn.Conv2d / torch.nn.functional.conv2d	INT	INT8, INT16, INT4	8x8, 8x4, 16x8
torch.nn.ELU / torch.nn.functional.elu	INT	INT8	8x8
torch.nn.functional.pad	INT	INT8	8x8
torch.nn.GELU / torch.nn.functional.gelu	INT	INT8, INT16	8x8, 16x8
torch.nn.Hardsigmoid / torch.nn.functional.hardsigmoid	INT	INT8	8x8
torch.nn.Hardswish / torch.nn.functional.hardswish	INT	INT8	8x8
torch.nn.Hardtanh / torch.nn.functional.hardtanh	INT	INT8	8x8
torch.nn.LeakyReLU / torch.nn.functional.leaky_relu	INT	INT8	8x8
torch.nn.Linear / torch.nn.functional.linear	INT	INT8, INT16	8x8, 16x8
torch.nn.SiLU / torch.nn.functional.silu	INT	INT8	8x8
torch.ones	INT	INT8	8x8
torch.permute / torch.Tensor.permute	INT	INT8, INT16	8x8, 16x8
torch.pixel_shuffle	INT	INT8	8x8
torch.pixel_unshuffle	INT	INT8	8x8
torch.pow / **	INT	INT8	8x8
torch.prelu	INT	INT8	8x8
torch.reciprocal	INT	INT8, INT16	8x8, 16x8
torch.relu / torch.nn.ReLU	INT	INT8	8x8
torch.remainder	INT	INT8	8x8
torch.repeat_interleave	INT	INT8, INT16	8x8, 16x8
torch.round	INT	INT8	8x8
torch.rsqrt	INT	INT8, INT16	8x8, 16x8
torch.scalar_tensor	INT	INT8	8x8
torch.select / torch.Tensor.select	INT	INT8	8x8
torch.sigmoid / torch.nn.Sigmoid	INT	INT8, INT16	8x8, 16x8
torch.sign	INT	INT8	8x8
torch.sin	INT	INT8	8x8
torch.sinh	INT	INT8	8x8
torch.softmax	INT	INT8	8x8
torch.split / torch.Tensor.split	INT	INT8, INT16	8x8, 16x8
torch.squeeze	INT	INT8	8x8
torch.squeeze / torch.Tensor.squeeze	INT	INT8	8x8
torch.stack	INT	INT8	8x8
torch.sub / -	INT	INT8, INT16	8x8, 16x8
torch.sum	INT	INT8	8x8
torch.t / torch.Tensor.t	INT	INT8	8x8
torch.tan	INT	INT8	8x8
torch.tanh / torch.nn.Tanh	INT	INT8, INT16	8x8, 16x8
torch.Tensor.__getitem__ / tensor indexing	INT	INT8	8x8
torch.Tensor.__getitem__ / tensor slicing	INT	INT8, INT16	8x8, 16x8
torch.Tensor.__setitem__ / tensor indexing assignment	INT	INT8	8x8
torch.Tensor.copy_	INT	INT8	8x8
torch.Tensor.expand	INT	INT8	8x8
torch.Tensor.repeat	INT	INT8, INT16	8x8, 16x8
torch.Tensor.unfold	INT	INT8	8x8
torch.Tensor.view	INT	INT8, INT16	8x8, 16x8
torch.transpose / torch.Tensor.transpose	INT	INT8	8x8
torch.unflatten	INT	INT8	8x8
torch.unsqueeze / torch.Tensor.unsqueeze	INT	INT8	8x8
torch.where	INT	INT8, INT16	8x8, 16x8
torch.while_loop	INT	INT8	8x8
torch.zeros	INT	INT8	8x8