Add split_with_sizes_copy (#537)

Summary:
Pull Request resolved: #537

Adds an implementation of `aten.split_with_sizes_copy`.

The implementation is identical to the existing implementation for `split.Tensor`. The only difference is the input checking and output resizing functions.

Reviewed By: manuelcandales

Differential Revision: D49763676

fbshipit-source-id: a1b58d6d78e3fc80ab988d21c80756bf3263b6b4

Author: SS-JIA

exir/dialects/edge/op/sample_input.py

@@ -1150,6 +1150,23 @@
             ),
         ],
     },
+    "split_with_sizes_copy.default": {  # (Tensor self, SymInt[] split_sizes, int dim=0) -> Tensor[]
+        "args": [
+            InArg(ArgType.Tensor, size=[2, 6, 3]),
+            InArg(ArgType.LengthList, value=[3, 1, 2]),
+            InArg(ArgType.Dim, value=1),
+        ],
+        "returns": [
+            Return(
+                ArgType.TensorList,
+                value=[
+                    Return(ArgType.Tensor, size=[2, 3, 3]),
+                    Return(ArgType.Tensor, size=[2, 1, 3]),
+                    Return(ArgType.Tensor, size=[2, 2, 3]),
+                ],
+            ),
+        ],
+    },
     "sqrt.default": {  # (Tensor self) -> Tensor
         "args": [
             InArg(ArgType.Tensor),

kernels/portable/cpu/op_split_with_sizes_copy.cpp

@@ -0,0 +1,87 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under the BSD-style license found in the
+ * LICENSE file in the root directory of this source tree.
+ */
+
+#include <cstdint>
+#include <cstring>
+
+#include <executorch/kernels/portable/cpu/util/copy_ops_util.h>
+#include <executorch/runtime/kernel/kernel_includes.h>
+
+namespace torch {
+namespace executor {
+namespace native {
+
+using Tensor = exec_aten::Tensor;
+using TensorList = exec_aten::TensorList;
+
+void split_with_sizes_copy_out(
+    RuntimeContext& ctx,
+    const Tensor& in,
+    exec_aten::ArrayRef<int64_t> split_sizes,
+    int64_t dim,
+    TensorList out) {
+  (void)ctx;
+  // Support python-style negative indexing. Note that this op does not accept 0
+  // dimensional input tensors.
+  if (dim < 0) {
+    dim += in.dim();
+  }
+
+  ET_KERNEL_CHECK(
+      ctx,
+      check_split_with_sizes_copy_args(in, split_sizes, dim, out),
+      InvalidArgument,
+      out);
+
+  Tensor::SizesType expected_out_size[kTensorDimensionLimit];
+  size_t expected_out_dim = 0;
+  for (size_t i = 0; i < split_sizes.size(); i++) {
+    expected_out_size[expected_out_dim++] = split_sizes[i];
+    get_split_with_sizes_copy_out_target_size(
+        in, split_sizes[i], dim, expected_out_size, &expected_out_dim);
+    ET_KERNEL_CHECK(
+        ctx,
+        resize_tensor(out[i], {expected_out_size, expected_out_dim}) ==
+            Error::Ok,
+        InvalidArgument,
+        out);
+  }
+
+  const size_t leading_dims = getLeadingDims(in, dim);
+  const size_t trailing_dims = getTrailingDims(in, dim);
+  const size_t step = in.size(dim) * trailing_dims;
+
+  ScalarType in_type = in.scalar_type();
+  ScalarType out_type = out[0].scalar_type();
+
+  ET_SWITCH_REAL_TYPES_AND(Bool, in_type, ctx, __func__, CTYPE_IN, [&]() {
+    ET_SWITCH_REAL_TYPES_AND(Bool, out_type, ctx, __func__, CTYPE_OUT, [&]() {
+      const CTYPE_IN* in_data = in.const_data_ptr<CTYPE_IN>();
+      for (size_t i = 0, e = out.size(); i < e; ++i) {
+        size_t out_step = out[i].size(dim) * trailing_dims;
+        if (out_step == 0) {
+          continue;
+        }
+        const CTYPE_IN* src = in_data;
+        CTYPE_OUT* dest = out[i].mutable_data_ptr<CTYPE_OUT>();
+        for (size_t j = 0; j < leading_dims; ++j) {
+          for (size_t k = 0; k < out_step; ++k) {
+            dest[k] = convert<CTYPE_OUT, CTYPE_IN>(src[k]);
+          }
+          src += step;
+          dest += out_step;
+        }
+        in_data += out_step;
+      }
+    });
+  });
+}
+
+} // namespace native
+} // namespace executor
+} // namespace torch

kernels/portable/cpu/targets.bzl

@@ -691,6 +691,12 @@ _ATEN_OPS = (
     op_target(
         name = "op_split_copy",
     ),
+    op_target(
+        name = "op_split_with_sizes_copy",
+        deps = [
+            "//executorch/kernels/portable/cpu/util:copy_ops_util",
+        ],
+    ),
     op_target(
         name = "op_sqrt",
         deps = [

kernels/portable/cpu/util/copy_ops_util.cpp

@@ -163,6 +163,47 @@ void get_pixel_shuffle_out_target_size(
   out_sizes[i] = in.size(i) * casted_upscale_factor;
 }
 
+bool check_split_with_sizes_copy_args(
+    const Tensor& in,
+    exec_aten::ArrayRef<int64_t> split_sizes,
+    int64_t dim,
+    TensorList out) {
+  ET_LOG_AND_RETURN_IF_FALSE(tensor_has_rank_greater_or_equal_to(in, 1));
+  ET_LOG_AND_RETURN_IF_FALSE(tensor_has_dim(in, dim));
+
+  ET_LOG_MSG_AND_RETURN_IF_FALSE(
+      split_sizes.size() == out.size(),
+      "Number of split sizes must match the number of output tensors");
+
+  int64_t sum = 0;
+  for (int i = 0; i < split_sizes.size(); i++) {
+    ET_LOG_MSG_AND_RETURN_IF_FALSE(
+        split_sizes[i] >= 0, "All split sizes must be non negative.");
+    sum += split_sizes[i];
+  }
+
+  const ssize_t dim_size = in.size(dim);
+  ET_LOG_MSG_AND_RETURN_IF_FALSE(
+      sum == dim_size,
+      "Sum of split sizes does not match input size at given dim");
+
+  return true;
+}
+
+void get_split_with_sizes_copy_out_target_size(
+    const Tensor& in,
+    int64_t split_size,
+    int64_t dim,
+    Tensor::SizesType* out_sizes,
+    size_t* out_ndim) {
+  *out_ndim = in.dim();
+
+  for (size_t d = 0; d < in.dim(); ++d) {
+    out_sizes[d] = in.size(d);
+  }
+  out_sizes[dim] = split_size;
+}
+
 bool check_stack_args(
     exec_aten::ArrayRef<Tensor> tensors,
     int64_t dim,

kernels/portable/cpu/util/copy_ops_util.h

@@ -43,6 +43,19 @@ void get_pixel_shuffle_out_target_size(
     Tensor::SizesType* out_sizes,
     size_t* out_ndim);
 
+bool check_split_with_sizes_copy_args(
+    const Tensor& in,
+    exec_aten::ArrayRef<int64_t> split_sizes,
+    int64_t dim,
+    TensorList out);
+
+void get_split_with_sizes_copy_out_target_size(
+    const Tensor& in,
+    int64_t split_size,
+    int64_t dim,
+    Tensor::SizesType* out_sizes,
+    size_t* out_ndim);
+
 bool check_stack_args(
     exec_aten::ArrayRef<Tensor> tensors,
     int64_t dim,

kernels/portable/functions.yaml

@@ -632,6 +632,11 @@
     - arg_meta: null
       kernel_name: torch::executor::split_copy_Tensor_out
 
+- op: split_with_sizes_copy.out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::split_with_sizes_copy_out
+
 - op: sqrt.out
   kernels:
     - arg_meta: null

kernels/test/op_split_with_sizes_copy_test.cpp

@@ -0,0 +1,90 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under the BSD-style license found in the
+ * LICENSE file in the root directory of this source tree.
+ */
+
+#include <executorch/kernels/test/FunctionHeaderWrapper.h> // Declares the operator
+#include <executorch/kernels/test/TestUtil.h>
+#include <executorch/kernels/test/supported_features.h>
+#include <executorch/runtime/core/exec_aten/exec_aten.h>
+#include <executorch/runtime/core/exec_aten/testing_util/tensor_factory.h>
+#include <executorch/runtime/core/exec_aten/testing_util/tensor_util.h>
+
+#include <gtest/gtest.h>
+
+using namespace ::testing;
+
+void op_split_with_sizes_copy_out(
+    const exec_aten::Tensor& self,
+    exec_aten::ArrayRef<int64_t> split_sizes,
+    int64_t dim,
+    exec_aten::TensorList out) {
+  exec_aten::RuntimeContext context{};
+  return torch::executor::aten::split_with_sizes_copy_outf(
+      context, self, split_sizes, dim, out);
+}
+
+TEST(OpSplitWithSizesCopyOutTest, SanityCheckDim1) {
+  torch::executor::testing::TensorFactory<exec_aten::ScalarType::Float> tfFloat;
+
+  exec_aten::Tensor self = tfFloat.make(
+      {2, 6, 3},
+      {-31.25,  -92.75,  -39.75,  -3.25,   53.875,  88.25,   -0.625,  -1.125,
+       14.75,   42.0,    89.875,  -21.125, -8.0,    -64.125, 23.0,    37.0,
+       46.125,  -83.25,  -58.125, 19.625,  -71.125, 64.75,   -1.375,  -83.5,
+       -61.375, 13.125,  28.625,  -94.0,   -67.0,   -8.625,  -88.875, -79.125,
+       0.375,   -61.375, 65.0,    -99.375});
+  ::std::vector<int64_t> split_sizes_vec = {3, 1, 2};
+  exec_aten::ArrayRef<int64_t> split_sizes = exec_aten::ArrayRef<int64_t>(
+      split_sizes_vec.data(), split_sizes_vec.size());
+  int64_t dim = 1;
+  ::std::vector<exec_aten::Tensor> out_vec = {
+      tfFloat.zeros({2, 3, 3}),
+      tfFloat.zeros({2, 1, 3}),
+      tfFloat.zeros({2, 2, 3})};
+  exec_aten::TensorList out =
+      exec_aten::TensorList(out_vec.data(), out_vec.size());
+  ::std::vector<exec_aten::Tensor> out_expected_vec = {
+      tfFloat.make(
+          {2, 3, 3},
+          {-31.25,
+           -92.75,
+           -39.75,
+           -3.25,
+           53.875,
+           88.25,
+           -0.625,
+           -1.125,
+           14.75,
+           -58.125,
+           19.625,
+           -71.125,
+           64.75,
+           -1.375,
+           -83.5,
+           -61.375,
+           13.125,
+           28.625}),
+      tfFloat.make({2, 1, 3}, {42.0, 89.875, -21.125, -94.0, -67.0, -8.625}),
+      tfFloat.make(
+          {2, 2, 3},
+          {-8.0,
+           -64.125,
+           23.0,
+           37.0,
+           46.125,
+           -83.25,
+           -88.875,
+           -79.125,
+           0.375,
+           -61.375,
+           65.0,
+           -99.375})};
+  exec_aten::TensorList out_expected =
+      exec_aten::TensorList(out_expected_vec.data(), out_expected_vec.size());
+  op_split_with_sizes_copy_out(self, split_sizes, dim, out);
+  EXPECT_TENSOR_LISTS_CLOSE(out, out_expected);
+}

kernels/test/targets.bzl

@@ -260,6 +260,7 @@ def define_common_targets():
     _common_op_test("op_slice_copy_test", ["aten", "portable"])
     _common_op_test("op_softmax_test", ["aten", "portable"])
     _common_op_test("op_split_copy_test", ["aten", "portable"])
+    _common_op_test("op_split_with_sizes_copy_test", ["aten", "portable"])
     _common_op_test("op_sqrt_test", ["aten", "portable"])
     _common_op_test("op_squeeze_copy_test", ["aten", "portable"])
     _common_op_test("op_stack_test", ["aten", "portable"])