Optimize index_out via fast path

Differential Revision: D81142086

Pull Request resolved: pytorch#13731

Author: manuelcandales

kernels/portable/cpu/op_index.cpp

@@ -22,21 +22,189 @@ namespace native {
 using Tensor = executorch::aten::Tensor;
 using TensorOptList = executorch::aten::ArrayRef<std::optional<Tensor>>;
 
-Tensor& index_Tensor_out(
+namespace {
+
+bool check_fast_path_conditions(
+    ET_UNUSED const Tensor& in,
+    TensorOptList indices,
+    size_t* dim) {
+  bool found_index = false;
+  for (const auto i : c10::irange(indices.size())) {
+    if (indices[i].has_value()) {
+      *dim = i;
+      // Fast path only supports a single non-null index tensor
+      if (found_index) {
+        return false;
+      }
+      found_index = true;
+      const Tensor& index = indices[i].value();
+      ScalarType ix_type = index.scalar_type();
+      // Fast path only supports Long or Int index tensors
+      if (ix_type != ScalarType::Long && ix_type != ScalarType::Int) {
+        return false;
+      }
+      // Fast path only supports a 1-dimensional index tensor
+      if (index.dim() != 1) {
+        return false;
+      }
+    }
+  }
+
+  // Fast path needs at least one non-null index tensor
+  if (!found_index) {
+    return false;
+  }
+
+  return true;
+}
+
+bool check_fast_path_args(
     KernelRuntimeContext& ctx,
     const Tensor& in,
     TensorOptList indices,
+    size_t dim,
     Tensor& out) {
-  (void)ctx;
+  ET_LOG_AND_RETURN_IF_FALSE(tensors_have_same_dtype(in, out));
+
+  ET_CHECK_OR_RETURN_FALSE(
+      static_cast<ssize_t>(indices.size()) <= in.dim(),
+      "Indexing too many dimensions");
+
+  const Tensor& index = indices[dim].value();
+
+  bool is_valid_index = true;
+  ET_SWITCH_TWO_TYPES(
+      Long, Int, index.scalar_type(), ctx, "index.Tensor", CTYPE, [&]() {
+        const CTYPE* const index_arr = index.const_data_ptr<CTYPE>();
+        for (const auto i : c10::irange(index.numel())) {
+          if (index_arr[i] < 0 ||
+              index_arr[i] >= static_cast<CTYPE>(in.size(dim))) {
+            ET_LOG(
+                Error,
+                "Index %" PRId64
+                " out of range for tensor with size %zd"
+                " at dimension %zu",
+                static_cast<int64_t>(index_arr[i]),
+                in.size(dim),
+                dim);
+            is_valid_index = false;
+            break;
+          }
+        }
+      });
+
+  ET_CHECK_OR_RETURN_FALSE(
+      is_valid_index,
+      "Some index values are not within bounds of input tensor at indexed dim");
 
+  return true;
+}
+
+void get_fast_path_index_out_target_size(
+    const Tensor& in,
+    TensorOptList indices,
+    size_t dim,
+    Tensor::SizesType* out_sizes,
+    size_t* out_ndim) {
+  *out_ndim = in.dim();
+
+  for (const auto d : c10::irange(static_cast<size_t>(in.dim()))) {
+    if (d != dim) {
+      out_sizes[d] = static_cast<Tensor::SizesType>(in.size(d));
+    } else {
+      out_sizes[d] =
+          static_cast<Tensor::SizesType>(indices[dim].value().numel());
+    }
+  }
+}
+
+Tensor& fast_path(
+    KernelRuntimeContext& ctx,
+    const Tensor& in,
+    TensorOptList indices,
+    size_t dim,
+    Tensor& out) {
   ET_KERNEL_CHECK(
-      ctx, check_index_args(in, indices, out), InvalidArgument, out);
+      ctx,
+      check_fast_path_args(ctx, in, indices, dim, out),
+      InvalidArgument,
+      out);
+
+  const Tensor& index = indices[dim].value();
+  ScalarType index_type = index.scalar_type();
+
+  // @lint-ignore CLANGTIDY facebook-hte-CArray
+  Tensor::SizesType expected_size[kTensorDimensionLimit];
+  size_t expected_ndim = 0;
+  get_fast_path_index_out_target_size(
+      in, indices, dim, expected_size, &expected_ndim);
 
+  ET_KERNEL_CHECK(
+      ctx,
+      resize_tensor(out, {expected_size, expected_ndim}) == Error::Ok,
+      InvalidArgument,
+      out);
+
+  if (out.dim() == 0) {
+    memcpy(out.mutable_data_ptr(), in.const_data_ptr(), out.nbytes());
+    return out;
+  }
+
+  size_t leading_dims = getLeadingDims(in, dim);
+  size_t trailing_dims = getTrailingDims(in, dim);
+
+  if (leading_dims == 0 || trailing_dims == 0) {
+    return out;
+  }
+
+  size_t in_dim_length = in.size(dim);
+  size_t out_dim_length = out.size(dim);
+
+  size_t length_per_step = trailing_dims * in.element_size();
+
+  const char* in_data = in.const_data_ptr<char>();
+  char* out_data = out.mutable_data_ptr<char>();
+
+  // @lint-ignore CLANGTIDY facebook-hte-CArray
+  static constexpr const char op_name[] = "index.Tensor_out";
+
+  ET_SWITCH_TWO_TYPES(Long, Int, index_type, ctx, op_name, CTYPE, [&]() {
+    const CTYPE* const index_arr = index.const_data_ptr<CTYPE>();
+    for (const auto i : c10::irange(leading_dims)) {
+      const char* src = in_data + i * in_dim_length * length_per_step;
+      char* dest = out_data + i * out_dim_length * length_per_step;
+      for (const auto j : c10::irange(out_dim_length)) {
+        const char* copy_src = src + index_arr[j] * length_per_step;
+        char* copy_dest = dest + j * length_per_step;
+        memcpy(copy_dest, copy_src, length_per_step);
+      }
+    }
+  });
+
+  return out;
+}
+
+} // namespace
+
+Tensor& index_Tensor_out(
+    KernelRuntimeContext& ctx,
+    const Tensor& in,
+    TensorOptList indices,
+    Tensor& out) {
   ET_KERNEL_CHECK(
       ctx, tensors_have_same_dim_order(in, out), InvalidArgument, out);
 
   ET_KERNEL_CHECK(ctx, tensor_is_default_dim_order(in), InvalidArgument, out);
 
+  size_t dim = 0;
+  bool is_fast_path = check_fast_path_conditions(in, indices, &dim);
+  if (is_fast_path) {
+    return fast_path(ctx, in, indices, dim, out);
+  }
+
+  ET_KERNEL_CHECK(
+      ctx, check_index_args(in, indices, out), InvalidArgument, out);
+
   ScalarType in_type = in.scalar_type();
   size_t block_count = count_index_blocks(indices);