ai-edge-torch-nightly 0.3.0.dev20250114__py3-none-any.whl

ai_edge_torch/_convert/fx_passes/optimize_layout_transposes_pass/layout_rewrite.py

@@ -0,0 +1,449 @@
+# Copyright 2024 The AI Edge Torch Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Layout rewrite for the optimized layout transposes pass."""
+
+import operator
+
+import ai_edge_torch
+from ai_edge_torch._convert.fx_passes.optimize_layout_transposes_pass import layout_mark
+from ai_edge_torch._convert.fx_passes.optimize_layout_transposes_pass import op_func_registry
+from ai_edge_torch._convert.fx_passes.optimize_layout_transposes_pass import utils
+import torch
+import torch.utils._pytree as pytree
+
+aten = torch.ops.aten
+StableHLOCompositeBuilder = ai_edge_torch.hlfb.StableHLOCompositeBuilder
+
+__all__ = ["rewrite_nhwc_node", "has_nhwc_rewriter"]
+
+
+class NHWCNodeRewritersRegistry(op_func_registry.OpFuncRegistry):
+
+  def __missing__(self, op):
+    def _rewriter(node):
+      raise RuntimeError(f"NHWC node rewriter not found: {str(node)}")
+
+    return _rewriter
+
+
+rewriters = NHWCNodeRewritersRegistry()
+
+
+def rewrite_nhwc_node(node: torch.fx.Node):
+  if not layout_mark.is_nhwc_node(node):
+    return
+
+  rewriters[node.target](node)
+
+
+def has_nhwc_rewriter(node: torch.fx.Node):
+  return node.target in rewriters
+
+
+# ======= Quantize ops
+
+
+@rewriters.register(torch.ops.quantized_decomposed.dequantize_per_tensor)
+@rewriters.register(torch.ops.quantized_decomposed.quantize_per_tensor)
+def noop(node: torch.fx.Node):
+  pass
+
+
+@rewriters.register(torch.ops.quantized_decomposed.dequantize_per_channel)
+@rewriters.register(torch.ops.quantized_decomposed.quantize_per_channel)
+def _qdq_per_channel_rewriter(node: torch.fx.Node):
+  new_args = []
+  new_kwargs = {}
+
+  def axis_nchw_to_nhwc(axis: int):
+    axis = axis if axis >= 0 else 4 + axis
+    return {3: 2, 2: 1, 1: 3}.get(axis, axis)
+
+  for arg, spec in zip(node.args, op._schema.arguments):
+    if spec.name == "axis":
+      new_args.append(axis_nchw_to_nhwc(arg))
+    else:
+      new_args.append(arg)
+
+  for spec in op._schema.arguments[len(node.args) :]:
+    if spec.name not in node.kwargs:
+      continue
+
+    if spec.name == "axis":
+      new_kwargs[spec.name] = axis_nchw_to_nhwc(node.kwargs[spec.name])
+    else:
+      new_kwargs[spec.name] = node.kwargs[spec.name]
+
+  node.args = tuple(new_args)
+  node.kwargs = new_kwargs
+
+
+# ======= Noop ops (layout insensitive ops)
+
+
+@rewriters.register(utils.tensor_to_nhwc)
+@rewriters.register(utils.tensor_to_nchw)
+@rewriters.register(operator.getitem)
+@rewriters.register("output")
+@rewriters.register(aten.add.Tensor)
+@rewriters.register(aten.add.Scalar)
+@rewriters.register(aten.atan2.default)
+@rewriters.register(aten.atan2.out)
+@rewriters.register(aten.bitwise_and.Tensor)
+@rewriters.register(aten.bitwise_and.Scalar)
+@rewriters.register(aten.bitwise_or.Tensor)
+@rewriters.register(aten.bitwise_or.Scalar)
+@rewriters.register(aten.bitwise_xor.Tensor)
+@rewriters.register(aten.bitwise_xor.Scalar)
+@rewriters.register(aten.div.Tensor)
+@rewriters.register(aten.div.Scalar)
+@rewriters.register(aten.div.Tensor_mode)
+@rewriters.register(aten.div.Scalar_mode)
+@rewriters.register(aten.fmod.Tensor)
+@rewriters.register(aten.fmod.Scalar)
+@rewriters.register(aten.mul.Tensor)
+@rewriters.register(aten.mul.Scalar)
+@rewriters.register(aten.remainder.Tensor)
+@rewriters.register(aten.remainder.Scalar)
+@rewriters.register(aten.sub.Tensor)
+@rewriters.register(aten.sub.Scalar)
+@rewriters.register(aten.eq.Tensor)
+@rewriters.register(aten.eq.Scalar)
+@rewriters.register(aten.ne.Tensor)
+@rewriters.register(aten.ne.Scalar)
+@rewriters.register(aten.le.Tensor)
+@rewriters.register(aten.le.Scalar)
+@rewriters.register(aten.ge.Tensor)
+@rewriters.register(aten.ge.Scalar)
+@rewriters.register(aten.gt.Tensor)
+@rewriters.register(aten.gt.Scalar)
+@rewriters.register(aten.lt.Tensor)
+@rewriters.register(aten.lt.Scalar)
+@rewriters.register(aten.maximum.default)
+@rewriters.register(aten.minimum.default)
+@rewriters.register(aten.mean.default)
+@rewriters.register(aten.prod.default)
+@rewriters.register(aten.abs.default)
+@rewriters.register(aten.acos.default)
+@rewriters.register(aten.acosh.default)
+@rewriters.register(aten.asin.default)
+@rewriters.register(aten.asinh.default)
+@rewriters.register(aten.atan.default)
+@rewriters.register(aten.atanh.default)
+@rewriters.register(aten.bitwise_not.default)
+@rewriters.register(aten.ceil.default)
+@rewriters.register(aten.clamp.default)
+@rewriters.register(aten.clamp.Tensor)
+@rewriters.register(aten.cos.default)
+@rewriters.register(aten.cosh.default)
+@rewriters.register(aten.erf.default)
+@rewriters.register(aten.exp.default)
+@rewriters.register(aten.expm1.default)
+@rewriters.register(aten.floor.default)
+@rewriters.register(aten.log.default)
+@rewriters.register(aten.log10.default)
+@rewriters.register(aten.log1p.default)
+@rewriters.register(aten.log2.default)
+@rewriters.register(aten.isnan.default)
+@rewriters.register(aten.neg.default)
+@rewriters.register(aten.pow.Tensor_Tensor)
+@rewriters.register(aten.pow.Tensor_Scalar)
+@rewriters.register(aten.pow.Scalar)
+@rewriters.register(aten.reciprocal.default)
+@rewriters.register(aten.round.default)
+@rewriters.register(aten.rsqrt.default)
+@rewriters.register(aten.sigmoid.default)
+@rewriters.register(aten.sign.default)
+@rewriters.register(aten.sin.default)
+@rewriters.register(aten.sinh.default)
+@rewriters.register(aten.sqrt.default)
+@rewriters.register(aten.tan.default)
+@rewriters.register(aten.tanh.default)
+@rewriters.register(aten.trunc.default)
+@rewriters.register(aten.nonzero.default)
+@rewriters.register(aten.copy.default)
+@rewriters.register(aten.mm.default)
+@rewriters.register(aten.fill.Scalar)
+@rewriters.register(aten.col2im.default)
+@rewriters.register(aten.addmm.default)
+@rewriters.register(aten.gelu.default)
+@rewriters.register(aten.hardtanh.default)
+@rewriters.register(aten.leaky_relu.default)
+@rewriters.register(aten.relu.default)
+@rewriters.register(aten.arange.start_step)
+@rewriters.register(aten.isinf.default)
+@rewriters.register(aten.logical_and.default)
+@rewriters.register(aten.logical_not.default)
+@rewriters.register(aten.logical_or.default)
+@rewriters.register(aten.logical_xor.default)
+@rewriters.register(aten.where.self)
+@rewriters.register(aten.clone.default)
+@rewriters.register(aten.any.default)
+@rewriters.register(aten.repeat.default)
+@rewriters.register(aten.alias.default)
+@rewriters.register(aten._pdist_forward.default)
+@rewriters.register(aten._cdist_forward.default)
+@rewriters.register(aten.bmm.default)
+@rewriters.register(aten.hardswish)
+@rewriters.register(aten.hardsigmoid)
+@rewriters.register(aten._to_copy)
+@rewriters.register(aten._prelu_kernel)
+@rewriters.register(aten.softplus)
+@rewriters.register(aten.silu)
+def noop(node: torch.fx.Node):
+  pass
+
+
+# ======= Add transposes before and after NCHW-only ops (T-aten-T)
+
+
+@rewriters.register(aten.upsample_bilinear2d)
+@rewriters.register(aten.upsample_nearest2d)
+@rewriters.register(aten.max_pool2d)
+@rewriters.register(aten.max_pool2d_with_indices)
+@rewriters.register(aten.avg_pool2d)
+@rewriters.register(aten._adaptive_avg_pool2d.default)
+def transpose_first_arg_rewriter(node: torch.fx.Node):
+  op = node.target
+
+  def nhwc_op(x, *args, **kwargs):
+    nonlocal op
+    x = utils.tensor_to_nchw(x)
+    res = pytree.tree_map_only(
+        torch.Tensor,
+        utils.tensor_to_nhwc,
+        op(x, *args, **kwargs),
+    )
+    return res
+
+  node.target = nhwc_op
+
+
+@rewriters.register(aten.conv2d)
+@rewriters.register(aten.convolution)
+def _aten_convolution_rewriter(node: torch.fx.Node):
+  op = node.target
+
+  def conv_nhwc(input, weight, bias=None, *args, **kwargs):
+    nonlocal op
+    nhwc_bias = None
+    if bias is not None and len(bias.shape) == 1:
+      nhwc_bias = bias
+      bias = None
+
+    input = utils.tensor_to_nchw(input)
+    res = pytree.tree_map_only(
+        torch.Tensor,
+        utils.tensor_to_nhwc,
+        op(input, weight, bias, *args, **kwargs),
+    )
+
+    if nhwc_bias is not None:
+      res += nhwc_bias
+    return res
+
+  node.target = conv_nhwc
+
+
+# ======= Rewrite dim attribute(s)
+
+
+@rewriters.register(aten._softmax.default)
+@rewriters.register(aten.select.int)
+@rewriters.register(aten.slice.Tensor)
+@rewriters.register(aten.sum.dim_IntList)
+@rewriters.register(aten.mean.dim)
+@rewriters.register(aten.prod.dim_int)
+@rewriters.register(aten.var.dim)
+@rewriters.register(aten.var.correction)
+@rewriters.register(aten.slice_scatter.default)
+@rewriters.register(aten.diagonal.default)
+@rewriters.register(aten.select_scatter.default)
+@rewriters.register(aten.sym_size.int)
+@rewriters.register(aten.sym_stride.int)
+@rewriters.register(aten._log_softmax.default)
+@rewriters.register(aten.split_with_sizes.default)
+@rewriters.register(aten.squeeze.dim)
+@rewriters.register(aten.squeeze.dims)
+@rewriters.register(aten.scatter.value)
+@rewriters.register(aten.scatter.src)
+@rewriters.register(aten.scatter_add.default)
+@rewriters.register(aten.scatter_reduce.two)
+@rewriters.register(aten.any.dim)
+@rewriters.register(aten.any.dims)
+@rewriters.register(aten.flip.default)
+@rewriters.register(aten.index_select.default)
+@rewriters.register(aten.cumsum.default)
+@rewriters.register(aten.max.dim)
+@rewriters.register(aten.min.dim)
+@rewriters.register(aten.gather.default)
+@rewriters.register(aten.sort.default)
+@rewriters.register(aten.topk.default)
+@rewriters.register(aten.cat.default)
+def dim_attr_rewriter(node: torch.fx.Node):
+  op = node.target
+
+  new_args = []
+  new_kwargs = {}
+
+  def dims_nchw_to_nhwc(dims: list[int]):
+    def convert(dim: int):
+      dim = dim if dim >= 0 else 4 + dim
+      return {3: 2, 2: 1, 1: 3}.get(dim, dim)
+
+    dims = pytree.tree_map_only(int, convert, dims)
+    dims = pytree.tree_map_only(torch.SymInt, convert, dims)
+    return dims
+
+  for arg, spec in zip(node.args, op._schema.arguments):
+    if spec.name.startswith("dim"):
+      new_args.append(dims_nchw_to_nhwc(arg))
+    else:
+      new_args.append(arg)
+
+  for spec in op._schema.arguments[len(node.args) :]:
+    if spec.name not in node.kwargs:
+      continue
+
+    if spec.name.startswith("dim"):
+      new_kwargs[spec.name] = dims_nchw_to_nhwc(node.kwargs[spec.name])
+    else:
+      new_kwargs[spec.name] = node.kwargs[spec.name]
+
+  node.args = tuple(new_args)
+  node.kwargs = new_kwargs
+
+
+# ======= Others
+
+
+@rewriters.register(aten._native_batch_norm_legit_no_training.default)
+def _aten__native_batch_norm_legit_no_training(node):
+  def batch_norm(input, weight, bias, running_mean, running_var, momentum, eps):
+    a = input - running_mean
+    b = torch.sqrt(running_var + eps)
+    out = a / b
+    if weight is not None:
+      out = out * weight
+    if bias is not None:
+      out = out + bias
+    return out, None, None
+
+  node.target = batch_norm
+
+
+@rewriters.register(aten.group_norm.default)
+def _aten_group_norm(node):
+  def group_norm(input, num_groups: int, weight=None, bias=None, eps=1e-5):
+    is_composite_supported = (
+        ai_edge_torch.config.enable_group_norm_composite
+        and weight is not None
+        and bias is not None
+    )
+
+    builder = None
+    if is_composite_supported:
+      builder = StableHLOCompositeBuilder(
+          name="odml.group_norm",
+          attr={
+              "num_groups": num_groups,
+              "epsilon": eps,
+              "reduction_axes": [3],
+              "channel_axis": 3,
+          },
+      )
+      input, weight, bias = builder.mark_inputs(input, weight, bias)
+
+    input = utils.tensor_to_nchw(input)
+    output = aten.group_norm.default(input, num_groups, weight, bias, eps=eps)
+    output = utils.tensor_to_nhwc(output)
+
+    if builder is not None:
+      output = builder.mark_outputs(output)
+    return output
+
+  node.target = group_norm
+
+
+@rewriters.register(aten.native_group_norm.default)
+def _aten_native_group_norm(node):
+
+  def native_group_norm(
+      input,
+      weight,
+      bias,
+      batch_size: int,
+      num_channels: int,
+      flattened_inner_size: int,
+      num_groups: int,
+      eps: float,
+      **kwargs,
+  ):
+    input_reshaped = torch.reshape(
+        input,
+        [
+            batch_size,
+            flattened_inner_size,
+            num_groups,
+            num_channels // num_groups,
+        ],
+    )
+    reduction_dims = [1, 3]
+
+    biased_var, mean = torch.var_mean(
+        input_reshaped, dim=reduction_dims, unbiased=False, keepdim=True
+    )
+    rstd = torch.rsqrt(biased_var + eps)
+
+    out = (input_reshaped - mean) * rstd
+    out = torch.reshape(out, input.shape)
+
+    if weight is not None:
+      out = out * weight
+    if bias is not None:
+      out = out + bias
+
+    mean = torch.squeeze(mean, reduction_dims)
+    rstd = torch.squeeze(rstd, reduction_dims)
+
+    return out, mean, rstd
+
+  node.target = native_group_norm
+
+
+@rewriters.register(aten.index)
+@rewriters.register(aten._unsafe_index)
+def _aten_index(node):
+  op = node.target
+
+  def index_nhwc(x, indices=[], *args, **kwargs):
+    nonlocal op
+    indices = list(indices)
+    if len(indices) < 4:
+      indices += [None] * (4 - len(indices))
+
+    indices[1:4] = indices[2], indices[3], indices[1]
+    return op(x, indices, *args, **kwargs)
+
+  node.target = index_nhwc
+
+
+@rewriters.register(aten.reflection_pad2d.default)
+def _aten_reflection_pad2d(node):
+  def reflection_pad2d_nhwc(x, padding):
+    padding = [0, 0] + padding
+    return torch.nn.functional.pad(x, padding, mode="reflect")
+
+  node.target = reflection_pad2d_nhwc

ai_edge_torch/_convert/fx_passes/optimize_layout_transposes_pass/op_func_registry.py

@@ -0,0 +1,30 @@
+# Copyright 2024 The AI Edge Torch Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Op function registry for the optimized layout transposes pass."""
+
+from ai_edge_torch._convert.fx_passes.optimize_layout_transposes_pass import utils
+
+
+class OpFuncRegistry(dict):
+
+  def register(self, op):
+    ops = utils.flatten_torch_op_overloads(op)
+
+    def inner(func):
+      for op in ops:
+        self[op] = func
+      return func
+
+    return inner