Merge branch 'main' into google-to-main

.gitmodules

@@ -56,6 +56,3 @@
 [submodule "third_party/mlir-hlo"]
 	path = third_party/mlir-hlo
 	url = https://github.com/tensorflow/mlir-hlo.git
-[submodule "third_party/wasm-micro-runtime"]
-	path = third_party/wasm-micro-runtime
-	url = https://github.com/bytecodealliance/wasm-micro-runtime.git

CMakeLists.txt

@@ -385,7 +385,6 @@ add_subdirectory(build_tools/third_party/half EXCLUDE_FROM_ALL)
 add_subdirectory(build_tools/third_party/pffft EXCLUDE_FROM_ALL)
 add_subdirectory(build_tools/third_party/renderdoc_api EXCLUDE_FROM_ALL)
 add_subdirectory(build_tools/third_party/vulkan_memory_allocator EXCLUDE_FROM_ALL)
-add_subdirectory(build_tools/third_party/wasm-micro-runtime EXCLUDE_FROM_ALL)
 
 add_subdirectory(third_party/cpuinfo EXCLUDE_FROM_ALL)
 add_subdirectory(third_party/googletest EXCLUDE_FROM_ALL)

SUBMODULE_VERSIONS.txt

@@ -14,7 +14,6 @@ d8c7ee00a687ac369e62e2032514a93a9b413502 third_party/pybind11
 685f86471e9d26b3eb7676695a2e2cefb4551ae9 third_party/spirv_cross
 f8bf11a0253a32375c32cad92c841237b96696c0 third_party/spirv_headers
 a8fbda3000692e62b477383e81a5386f8305d0c6 third_party/tensorflow
-8732f0e94e4e41049a43029202bda94d7b4e85da third_party/tracy
+e35d02186d00eff26ec6c698331a0767a28a51b0 third_party/tracy
 9bd3f561bcee3f01d22912de10bb07ce4e23d378 third_party/vulkan_headers
 3528e2aed3e8808f33e1e7d63eeb1560456a605a third_party/vulkan_memory_allocator
-54e82ec439552018d0f26c60b0fd6cc9e36b42a8 third_party/wasm-micro-runtime

build_tools/bazel/build_core.sh

@@ -48,6 +48,9 @@ declare -a test_env_args=(
 declare -a default_build_tag_filters=("-nokokoro")
 declare -a default_test_tag_filters=("-nokokoro" "-driver=metal")
 
+# The VK_KHR_shader_float16_int8 extension is optional prior to Vulkan 1.2.
+# We test on SwiftShader, which does not support this extension.
+default_test_tag_filters+=("-vulkan_uses_vk_khr_shader_float16_int8")
 # CUDA CI testing disabled until we setup a target for it.
 default_test_tag_filters+=("-driver=cuda")
 

build_tools/buildkite/cmake/android/arm64-v8a/pipeline.yml

@@ -59,7 +59,9 @@ steps:
       - "tar xzf build-artifacts.tgz"
       - "find build-android/ -name '*.cmake' -exec sed -i \"s!\\$IREE_DOCKER_WORKDIR/!\\$PWD/!g\" {} \\;"
       - "cd build-android/"
-      - "ctest --output-on-failure"
+      # vulkan tests using khr_shader_float16_int8 are failing on pixel4.
+      # Disabling it until we identify the root cause.
+      - "ctest --output-on-failure --label-exclude \"^vulkan_uses_vk_khr_shader_float16_int8\\$\""
     agents:
       - "android-soc=snapdragon-855"
       - "queue=test-android"

build_tools/cmake/test.sh

@@ -30,6 +30,9 @@ export IREE_VULKAN_DISABLE=${IREE_VULKAN_DISABLE:-0}
 export IREE_LLVMAOT_DISABLE=${IREE_LLVMAOT_DISABLE:-0}
 # CUDA is off by default.
 export IREE_CUDA_DISABLE=${IREE_CUDA_DISABLE:-1}
+# The VK_KHR_shader_float16_int8 extension is optional prior to Vulkan 1.2.
+# We test on SwiftShader, which does not support this extension.
+export IREE_VULKAN_F16_DISABLE=${IREE_VULKAN_F16_DISABLE:-1}
 
 # Tests to exclude by label. In addition to any custom labels (which are carried
 # over from Bazel tags), every test should be labeled with the directory it is
@@ -62,6 +65,9 @@ fi
 if [[ "${IREE_CUDA_DISABLE?}" == 1 ]]; then
   label_exclude_args+=("^driver=cuda$")
 fi
+if [[ "${IREE_VULKAN_F16_DISABLE?}" == 1 ]]; then
+  label_exclude_args+=("^vulkan_uses_vk_khr_shader_float16_int8$")
+fi
 
 # Join on "|"
 label_exclude_regex="($(IFS="|" ; echo "${label_exclude_args[*]?}"))"

build_tools/kokoro/gcp_ubuntu/cmake-bazel/linux/x86-swiftshader/build.sh

@@ -71,4 +71,4 @@ export CTEST_PARALLEL_LEVEL=${CTEST_PARALLEL_LEVEL:-$(nproc)}
 echo "Testing with CTest"
 ctest --output-on-failure \
    --tests-regex "^integrations/tensorflow/|^bindings/python/" \
-   --label-exclude "^nokokoro$"
+   --label-exclude "^nokokoro$|^vulkan_uses_vk_khr_shader_float16_int8$"

build_tools/kokoro/gcp_ubuntu/cmake/linux/x86-turing/build.sh

@@ -44,6 +44,7 @@ echo "Initializing submodules"
 echo "Building with cmake"
 ./build_tools/cmake/clean_build.sh
 
+export IREE_VULKAN_F16_DISABLE=0
 export IREE_CUDA_DISABLE=0
 echo "Testing with ctest"
 ./build_tools/cmake/test.sh

build_tools/mako/config/benchmark_template.config

@@ -0,0 +1,54 @@
+# Copyright 2021 Google LLC
+#
+# 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
+#
+#      https://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.
+
+# This template is used to define/update Mako benchmarks. The input_value_info
+# and metric_info_list should be updated according to the purpose of the
+# benchmark. The benchmark_name and description must be updated. The
+# benchmark_key should not be added before creating the benchmark. You have to
+# fill the server-generated key which is returned by the Mako tool.
+#
+# For more details, see:
+#   https://github.com/google/mako/blob/master/docs/GUIDE.md#preparing-your-benchmark
+
+# benchmark_key: "TBD"
+benchmark_name: "BENCHMARK_NAME (PHONE) (UNIT)"
+project_name: "iree"
+
+# Only owners can write to the benchmark
+owner_list: "[email protected]"
+owner_list: "[email protected]"
+
+# Define the name and type for x-axis of run charts
+input_value_info: {
+  value_key: "t"
+  label: "time"
+  type: NUMERIC
+}
+
+# Three metrics, define the names for y-axis values of both run and aggregate
+# charts.
+metric_info_list: {
+  value_key: "cpu"
+  label: "DYLib_AOT"
+}
+metric_info_list: {
+  value_key: "vmla"
+  label: "VMLA"
+}
+metric_info_list: {
+  value_key: "vlk"
+  label: "Vulkan-SPIRV"
+}
+
+description: "IREE Benchmark Template"

build_tools/mako/config/mobilenet-v2-pixel4.config

@@ -41,5 +41,9 @@ metric_info_list: {
   value_key: "vlk"
   label: "Vulkan-SPIRV"
 }
+metric_info_list: {
+  value_key: "cpu2"
+  label: "DYLib_AOT_Linalg-On-Tensors"
+}
 
 description: "Keras Applications: MobileNet V2"

build_tools/mako/config/mobilenet-v2-s20.config

@@ -41,5 +41,13 @@ metric_info_list: {
   value_key: "vlk"
   label: "Vulkan-SPIRV"
 }
+metric_info_list: {
+  value_key: "cpu2"
+  label: "DYLib_AOT_Linalg-On-Tensors"
+}
+metric_info_list: {
+  value_key: "vlk2"
+  label: "Vulkan-SPIRV_Linalg-On-Tensors"
+}
 
 description: "Keras Applications: MobileNet V2"

build_tools/mako/configuration.py

@@ -94,7 +94,9 @@ def __init__(self, name, model_artifacts_name, model_path, flagfile_path,
     self.phones = phones
 
 
-def get_pixel4_default_target_list(batch_config=None):
+def get_pixel4_default_target_list(skipped_target=None, batch_config=None):
+  if skipped_target is None:
+    skipped_target = []
   if batch_config is None:
     batch_config = []
   targets = [
@@ -111,15 +113,36 @@ def get_pixel4_default_target_list(batch_config=None):
           compilation_flags=[
               "--iree-spirv-enable-vectorization",
               "--iree-vulkan-target-triple=qualcomm-adreno640-unknown-android10"
+          ]),
+      TargetInfo(name="dylib-llvm-aot",
+                 mako_tag="cpu2",
+                 compilation_flags=[
+                     "--iree-llvm-target-triple=aarch64-none-linux-android29",
+                     "--iree-flow-dispatch-linalg-on-tensors",
+                     "--iree-codegen-llvm-experimental-linalg-on-tensors",
+                     "-iree-flow-inline-constants-max-byte-length=2048"
+                 ],
+                 runtime_flags=["--dylib_worker_count=1"]),
+      TargetInfo(
+          name="vulkan-spirv",
+          mako_tag="vlk2",
+          compilation_flags=[
+              "--iree-vulkan-target-triple=qualcomm-adreno640-unknown-android10",
+              "--iree-flow-dispatch-linalg-on-tensors",
+              "--iree-codegen-spirv-experimental-linalg-on-tensors",
+              "-iree-flow-inline-constants-max-byte-length=2048"
           ])
   ]
+  targets = [elem for elem in targets if elem.mako_tag not in skipped_target]
   for target in targets:
     if target.mako_tag in batch_config:
       target.add_batch_flag(batch_config[target.mako_tag])
   return targets
 
 
-def get_s20_default_target_list(batch_config=None):
+def get_s20_default_target_list(skipped_target=None, batch_config=None):
+  if skipped_target is None:
+    skipped_target = []
   if batch_config is None:
     batch_config = []
   targets = [
@@ -130,13 +153,33 @@ def get_s20_default_target_list(batch_config=None):
                      "--iree-llvm-target-triple=aarch64-none-linux-android29"
                  ],
                  runtime_flags=["--dylib_worker_count=1"]),
-      TargetInfo(name="vulkan-spirv",
-                 mako_tag="vlk",
+      TargetInfo(
+          name="vulkan-spirv",
+          mako_tag="vlk",
+          compilation_flags=[
+              "--iree-spirv-enable-vectorization",
+              "--iree-vulkan-target-triple=valhall-g77-unknown-android10",
+          ]),
+      TargetInfo(name="dylib-llvm-aot",
+                 mako_tag="cpu2",
                  compilation_flags=[
-                     "--iree-spirv-enable-vectorization",
-                     "--iree-vulkan-target-triple=valhall-g77-unknown-android10"
-                 ])
+                     "--iree-llvm-target-triple=aarch64-none-linux-android29",
+                     "--iree-flow-dispatch-linalg-on-tensors",
+                     "--iree-codegen-llvm-experimental-linalg-on-tensors",
+                     "-iree-flow-inline-constants-max-byte-length=2048"
+                 ],
+                 runtime_flags=["--dylib_worker_count=1"]),
+      TargetInfo(
+          name="vulkan-spirv",
+          mako_tag="vlk2",
+          compilation_flags=[
+              "--iree-vulkan-target-triple=valhall-g77-unknown-android10",
+              "--iree-flow-dispatch-linalg-on-tensors",
+              "--iree-codegen-spirv-experimental-linalg-on-tensors",
+              "-iree-flow-inline-constants-max-byte-length=2048"
+          ])
   ]
+  targets = [elem for elem in targets if elem.mako_tag not in skipped_target]
   for target in targets:
     if target.mako_tag in batch_config:
       target.add_batch_flag(batch_config[target.mako_tag])
@@ -159,13 +202,16 @@ def get_s20_default_target_list(batch_config=None):
             PhoneBenchmarkInfo(name="Pixel4",
                                benchmark_key="5538704950034432",
                                targets=get_pixel4_default_target_list(
-                                   {'cpu': 8})),
+                                   skipped_target=['cpu2', 'vlk2'],
+                                   batch_config={'cpu': 8})),
             PhoneBenchmarkInfo(name="S20",
                                benchmark_key="4699630718681088",
-                               targets=get_s20_default_target_list({
-                                   'cpu': 8,
-                                   'vlk': 16
-                               })),
+                               targets=get_s20_default_target_list(
+                                   skipped_target=['cpu2', 'vlk2'],
+                                   batch_config={
+                                       'cpu': 8,
+                                       'vlk': 16
+                                   })),
         ]),
     ModelBenchmarkInfo(
         name="mobilenet-v2",
@@ -176,13 +222,20 @@ def get_s20_default_target_list(batch_config=None):
             PhoneBenchmarkInfo(name="Pixel4",
                                benchmark_key="6338759231537152",
                                targets=get_pixel4_default_target_list(
-                                   {'cpu': 16})),
-            PhoneBenchmarkInfo(name="S20",
-                               benchmark_key="5618403088793600",
-                               targets=get_s20_default_target_list({
-                                   'cpu': 16,
-                                   'vlk': 64
-                               })),
+                                   skipped_target=['vlk2'],
+                                   batch_config={
+                                       'cpu': 16,
+                                       'cpu2': 10
+                                   })),
+            PhoneBenchmarkInfo(
+                name="S20",
+                benchmark_key="5618403088793600",
+                targets=get_s20_default_target_list(batch_config={
+                    'cpu': 16,
+                    'vlk': 64,
+                    'cpu2': 10,
+                    'vlk2': 64
+                })),
         ])
 ]
 

docs/get_started/getting_started_tensorflow.md

@@ -39,6 +39,7 @@ For example to run TensorFlow-based tests, you can build `iree-import-tf`
 TODO(4979): Rename iree-tf-import to iree-import-tf
 
 ```shell
+python3 configure_bazel.py
 cd integrations/tensorflow
 bazel build \
   //iree_tf_compiler:iree-tf-import \