[pull] master from tensorflow:master

tensorflow/core/kernels/image/BUILD

@@ -264,6 +264,8 @@ tf_kernel_library(
     prefix = "extract_image_patches_op",
     deps = [
         "//tensorflow/core/kernels:ops_util",
+        "@com_google_absl//absl/status",
+        "@com_google_absl//absl/strings",
     ] + IMAGE_DEPS,
 )
 
@@ -309,6 +311,7 @@ tf_kernel_library(
     deps = IMAGE_DEPS + [
         ":sampling_kernels",
         "@com_google_absl//absl/status",
+        "@com_google_absl//absl/strings",
         "@com_google_absl//absl/strings:str_format",
     ],
 )

tensorflow/core/kernels/image/extract_image_patches_op.cc

@@ -16,6 +16,10 @@ limitations under the License.
 // See docs in ../ops/image_ops.cc.
 
 #include <cstdint>
+#include <string>
+
+#include "absl/status/status.h"
+#include "absl/strings/str_cat.h"
 #define USE_EIGEN_TENSOR
 #define EIGEN_USE_THREADS
 

tensorflow/core/kernels/image/scale_and_translate_op.cc

@@ -22,6 +22,7 @@ limitations under the License.
 #include <vector>
 
 #include "absl/status/status.h"
+#include "absl/strings/str_cat.h"
 #include "absl/strings/str_format.h"
 #define EIGEN_USE_THREADS
 

tensorflow/lite/delegates/ynnpack/BUILD

@@ -0,0 +1,47 @@
+load("@rules_cc//cc:cc_library.bzl", "cc_library")
+load("//tensorflow:tensorflow.default.bzl", "get_compatible_with_portable")
+load("//tensorflow/lite:build_def.bzl", "tflite_copts")
+
+package(
+    # copybara:uncomment default_applicable_licenses = ["//tensorflow:LICENSE"],
+    default_visibility = ["//visibility:public"],
+    licenses = ["notice"],
+)
+
+cc_library(
+    name = "ynnpack_delegate",
+    srcs = [
+        "copy.cc",
+        "copy.h",
+        "dot.cc",
+        "dot.h",
+        "elementwise.cc",
+        "elementwise.h",
+        "pooling.cc",
+        "pooling.h",
+        "reduction.cc",
+        "reduction.h",
+        "softmax.cc",
+        "softmax.h",
+        "utils.cc",
+        "utils.h",
+        "ynnpack_delegate.cc",
+    ],
+    hdrs = ["ynnpack_delegate.h"],
+    compatible_with = get_compatible_with_portable(),
+    copts = tflite_copts(),
+    linkstatic = True,
+    deps = [
+        "//tensorflow/lite:builtin_ops",
+        "//tensorflow/lite:minimal_logging",
+        "//tensorflow/lite/core/c:common",
+        "//tensorflow/lite/delegates/utils:simple_delegate",
+        "//tensorflow/lite/kernels:kernel_util",
+        "//tensorflow/lite/types:half",
+        "@XNNPACK//ynnpack",
+        "@XNNPACK//ynnpack/composites",
+        "@com_google_absl//absl/container:flat_hash_map",
+        "@com_google_absl//absl/types:span",
+        "@slinky//slinky/base:thread_pool_impl",
+    ],
+)

tensorflow/lite/delegates/ynnpack/README.md

@@ -0,0 +1,49 @@
+# YNNPACK Delegate for LiteRT
+
+> [!WARNING]
+> The YNNPACK delegate is **experimental** and under active development. Expect
+> bugs and performance issues when using it.
+
+The YNNPACK delegate allows LiteRT (formerly TensorFlow Lite) to offload
+supported operators to YNNPACK.
+
+YNNPACK aims to provide great flexibility with good performance.
+
+## Delegate Provider Options
+
+When using LiteRT tooling (e.g., benchmarks, evaluation tools) that link the
+`ynnpack_delegate_provider`, the following command-line flags are exposed to
+configure the YNNPACK delegate:
+
+### Core Options
+
+*   **`--use_ynnpack=true|false`** (default: `false`):
+    Explicitly apply the YNNPACK delegate to the model.
+
+*   **`--num_threads=N`** (default: `0` or `1` depending on tool):
+    The number of threads to use for execution. Note that YNNPACK will only use
+    a thread pool for `num_threads > 1`. A value of `0` or `1` disables the
+    thread pool (single-threaded execution).
+
+### YNNPACK Specific Options
+
+*   **`--ynnpack_static_shape=true|false`** (default: `false`):
+    Make input shapes static instead of dynamic. Enabling this may improve
+    execution (`Invoke`) performance by allowing YNNPACK to optimize for fixed
+    shapes, but it makes model reshaping (`ResizeInputTensor`) much more
+    expensive.
+
+*   **`--ynnpack_fast_math=true|false`** (default: `false`):
+    Enable `YNN_FLAG_FAST_MATH`. This allows YNNPACK to use faster but
+    potentially less precise mathematical approximations.
+
+*   **`--ynnpack_consistent_arithmetic=true|false`** (default: `false`):
+    Enable `YNN_FLAG_CONSISTENT_ARITHMETIC`. YNNPACK will attempt to provide
+    numerically consistent results for all hardware the **same build** of
+    YNNPACK runs on. It does not guarantee consistency across different builds
+    (which means it does not guarantee consistency across different platforms,
+    which are necessarily different builds).
+
+*   **`--ynnpack_no_excess_precision=true|false`** (default: `false`):
+    Enable `YNN_FLAG_NO_EXCESS_PRECISION`. YNNPACK will not promote tensors to
+    higher precision as a performance optimization.