[llm] Add arange() tensor maker API (pytorch#11965)

This PR was created by the merge bot to help merge the original PR into
the main branch.
ghstack PR number: pytorch#11861 by
@larryliu0820
^ Please use this as the source of truth for the PR details, comments,
and reviews
ghstack PR base:
https://github.com/pytorch/executorch/tree/gh/larryliu0820/66/base
ghstack PR head:
https://github.com/pytorch/executorch/tree/gh/larryliu0820/66/head
Merge bot PR base: https://github.com/pytorch/executorch/tree/main
Merge bot PR head:
https://github.com/pytorch/executorch/tree/gh/larryliu0820/66/orig
@diff-train-skip-merge

Co-authored-by: Mengwei Liu <[email protected]>

Author: pytorchbot

kernels/portable/cpu/op_arange.cpp

@@ -7,6 +7,7 @@
  */
 
 #include <executorch/kernels/portable/cpu/scalar_utils.h>
+#include <executorch/kernels/portable/cpu/util/arange_util.h>
 #include <executorch/kernels/portable/cpu/util/kernel_ops_util.h>
 #include <executorch/runtime/kernel/kernel_includes.h>
 #include <executorch/runtime/platform/assert.h>
@@ -29,22 +30,15 @@ Tensor& arange_out(KernelRuntimeContext& ctx, const Scalar& end, Tensor& out) {
 
   ET_KERNEL_CHECK(ctx, tensor_is_default_dim_order(out), InvalidArgument, out);
 
-  size_t size = static_cast<size_t>(std::ceil(end_val));
-
-  Tensor::SizesType out_length = static_cast<Tensor::SizesType>(size);
+  Tensor::SizesType out_length = compute_arange_out_size(0.0, end_val, 1.0);
 
   ET_KERNEL_CHECK(
       ctx,
       resize_tensor(out, {&out_length, 1}) == Error::Ok,
       InvalidArgument,
       out);
 
-  ET_SWITCH_REALHBF16_TYPES(out.scalar_type(), ctx, "arange.out", CTYPE, [&]() {
-    auto out_data = out.mutable_data_ptr<CTYPE>();
-    for (size_t i = 0; i < size; i++) {
-      out_data[i] = static_cast<CTYPE>(i);
-    }
-  });
+  arange_out_impl(ctx, end_val, out);
 
   return out;
 }
@@ -77,24 +71,16 @@ Tensor& arange_start_out(
 
   ET_KERNEL_CHECK(ctx, tensor_is_default_dim_order(out), InvalidArgument, out);
 
-  double size_d = (d_end - d_start) / d_step;
-  size_t size = static_cast<size_t>(std::ceil(size_d));
-
-  Tensor::SizesType out_length = static_cast<Tensor::SizesType>(size);
+  Tensor::SizesType out_length =
+      compute_arange_out_size(d_start, d_end, d_step);
 
   ET_KERNEL_CHECK(
       ctx,
       resize_tensor(out, {&out_length, 1}) == Error::Ok,
       InvalidArgument,
       out);
 
-  ET_SWITCH_REALHBF16_TYPES(
-      out.scalar_type(), ctx, "arange.start_out", CTYPE, [&]() {
-        auto out_data = out.mutable_data_ptr<CTYPE>();
-        for (size_t i = 0; i < size; i++) {
-          out_data[i] = convert<CTYPE, double>(d_start + i * d_step);
-        }
-      });
+  arange_out_impl(ctx, d_start, d_end, d_step, out);
 
   return out;
 }

kernels/portable/cpu/util/arange_util.cpp

@@ -0,0 +1,51 @@
+/*
+ * 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/portable/cpu/util/arange_util.h>
+
+namespace torch::executor::native {
+#define ET_ARANGE_IMPL(ctx, start, numel, step, out, op_name)               \
+  ET_SWITCH_REALHBF16_TYPES(out.scalar_type(), ctx, op_name, CTYPE, [&]() { \
+    auto out_data = out.mutable_data_ptr<CTYPE>();                          \
+    for (Tensor::SizesType i = 0; i < numel; ++i) {                         \
+      out_data[i] = static_cast<CTYPE>(start + i * step);                   \
+    }                                                                       \
+  })
+
+Tensor::SizesType
+compute_arange_out_size(double start, double end, double step) {
+  Tensor::SizesType numel =
+      static_cast<Tensor::SizesType>(std::ceil((end - start) / step));
+
+  ET_CHECK_MSG(
+      numel >= 0,
+      "numel should be non-negative, but got (%d). start (%f), end (%f), step (%f)",
+      numel,
+      start,
+      end,
+      step);
+  return numel;
+}
+
+void arange_out_impl(
+    KernelRuntimeContext& ctx,
+    double start,
+    double end,
+    double step,
+    Tensor& out) {
+  (void)ctx;
+  Tensor::SizesType numel = compute_arange_out_size(start, end, step);
+  ET_ARANGE_IMPL(ctx, start, numel, step, out, "arange.start_out");
+}
+
+void arange_out_impl(KernelRuntimeContext& ctx, double end, Tensor& out) {
+  (void)ctx;
+  ET_ARANGE_IMPL(ctx, 0.0, end, 1.0, out, "arange.out");
+}
+
+} // namespace torch::executor::native

kernels/portable/cpu/util/arange_util.h

@@ -0,0 +1,41 @@
+/*
+ * 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.
+ */
+
+#pragma once
+
+#include <executorch/runtime/kernel/kernel_includes.h>
+
+namespace torch::executor::native {
+
+Tensor::SizesType
+compute_arange_out_size(double start, double end, double step);
+
+inline Tensor::SizesType compute_arange_out_size(double end) {
+  return compute_arange_out_size(0.0, end, 1.0);
+}
+
+void arange_out_impl(
+    KernelRuntimeContext& ctx,
+    double start,
+    double end,
+    double step,
+    Tensor& out);
+
+void arange_out_impl(KernelRuntimeContext& ctx, double end, Tensor& out);
+
+inline void
+arange_out_impl(double start, double end, double step, Tensor& out) {
+  KernelRuntimeContext ctx;
+  arange_out_impl(ctx, start, end, step, out);
+}
+
+inline void arange_out_impl(double end, Tensor& out) {
+  KernelRuntimeContext ctx;
+  arange_out_impl(ctx, 0.0, end, 1.0, out);
+}
+} // namespace torch::executor::native

kernels/portable/cpu/util/targets.bzl

@@ -13,6 +13,7 @@ def define_common_targets():
         name = "all_deps",
         exported_deps = [
             "//executorch/extension/threadpool:threadpool",
+            "//executorch/kernels/portable/cpu/util:arange_util",
             "//executorch/kernels/portable/cpu/util:functional_util",
             "//executorch/kernels/portable/cpu/util:broadcast_util",
             "//executorch/kernels/portable/cpu/util:kernel_ops_util",
@@ -300,6 +301,19 @@ def define_common_targets():
         visibility = ["//executorch/kernels/portable/cpu/..."],
     )
 
+    runtime.cxx_library(
+        name = "arange_util",
+        srcs = ["arange_util.cpp"],
+        exported_headers = ["arange_util.h"],
+        deps = [
+            "//executorch/runtime/kernel:kernel_includes",
+        ],
+        visibility = [
+            "//executorch/kernels/portable/cpu/...",
+            "//executorch/extension/llm/...",
+        ],
+    )
+
     runtime.cxx_library(
         name = "broadcast_indexes_range",
         exported_headers = ["broadcast_indexes_range.h"],

shim_et/xplat/executorch/kernels/portable/op_registration_util.bzl

@@ -268,6 +268,7 @@ ATEN_OPS = (
     op_target(
         name = "op_arange",
         deps = [
+            "//executorch/kernels/portable/cpu/util:arange_util",
             "//executorch/kernels/portable/cpu/util:kernel_ops_util",
             "//executorch/runtime/core/exec_aten/util:scalar_type_util",
             "//executorch/runtime/core/exec_aten/util:tensor_util",