Release Frogfish

Author: dmitriykovalev

.gitignore

@@ -1,3 +1,2 @@
 bazel-*
 out/
-WORKSPACE

BUILD

@@ -25,6 +25,13 @@ exports_files([
     "LICENSE",
 ])
 
+config_setting(
+    name = "opt",
+    values = {
+        "compilation_mode": "opt",
+    },
+)
+
 # If --define darwinn_portable=1, compile without google3 deps.
 config_setting(
     name = "darwinn_portable",

Makefile

@@ -12,16 +12,15 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 SHELL := /bin/bash
+PYTHON3 ?= python3
 MAKEFILE_DIR := $(realpath $(dir $(lastword $(MAKEFILE_LIST))))
 OUT_DIR := $(MAKEFILE_DIR)/out
 OS := $(shell uname -s)
 
 ifeq ($(OS),Linux)
 CPU ?= k8
-WORKSPACE_PLATFORM_FILE := WORKSPACE.linux
 else ifeq ($(OS),Darwin)
 CPU ?= darwin
-WORKSPACE_PLATFORM_FILE := WORKSPACE.darwin
 else
 $(error $(OS) is not supported)
 endif
@@ -40,23 +39,16 @@ BAZEL_OUT_DIR := $(MAKEFILE_DIR)/bazel-out/$(CPU)-$(COMPILATION_MODE)/bin
 # Linux-specific parameters
 BAZEL_BUILD_TARGET_Linux := //tflite/public:libedgetpu_direct_all.so
 BAZEL_BUILD_FLAGS_Linux := --crosstool_top=@crosstool//:toolchains \
-                           --compiler=gcc \
-                           --linkopt=-l:libusb-1.0.so
+                           --compiler=gcc
 BAZEL_BUILD_OUTPUT_FILE_Linux := libedgetpu.so.1.0
 BAZEL_BUILD_OUTPUT_SYMLINK_Linux := libedgetpu.so.1
 
-ifeq ($(COMPILATION_MODE), opt)
-BAZEL_BUILD_FLAGS_Linux += --linkopt=-Wl,--strip-all
-endif
 ifeq ($(CPU), armv6)
 BAZEL_BUILD_FLAGS_Linux += --linkopt=-L/usr/lib/arm-linux-gnueabihf/
 endif
 
 # Darwin-specific parameters
 BAZEL_BUILD_TARGET_Darwin := //tflite/public:libedgetpu_direct_usb.dylib
-BAZEL_BUILD_FLAGS_Darwin := --linkopt=-L/opt/local/lib \
-                            --linkopt=-lusb-1.0 \
-                            --copt=-fvisibility=hidden
 BAZEL_BUILD_OUTPUT_FILE_Darwin := libedgetpu.1.0.dylib
 BAZEL_BUILD_OUTPUT_SYMLINK_Darwin := libedgetpu.1.dylib
 
@@ -65,11 +57,10 @@ BAZEL_BUILD_FLAGS := --sandbox_debug --subcommands \
   --experimental_repo_remote_exec \
   --compilation_mode=$(COMPILATION_MODE) \
   --define darwinn_portable=1 \
-  --copt=-DSTRIP_LOG=1 \
-  --copt=-fno-rtti \
-  --copt=-fno-exceptions \
-  --copt='-D__FILE__=""' \
-  --cpu=$(CPU)
+  --action_env PYTHON_BIN_PATH=$(shell which $(PYTHON3)) \
+  --cpu=$(CPU) \
+  --embed_label='TENSORFLOW_COMMIT=$(shell grep "TENSORFLOW_COMMIT =" $(MAKEFILE_DIR)/workspace.bzl | grep -o '[0-9a-f]\{40\}')' \
+  --stamp
 BAZEL_BUILD_FLAGS += $(BAZEL_BUILD_FLAGS_$(OS))
 BAZEL_BUILD_TARGET := $(BAZEL_BUILD_TARGET_$(OS))
 BAZEL_BUILD_OUTPUT_FILE := $(BAZEL_BUILD_OUTPUT_FILE_$(OS))
@@ -95,23 +86,31 @@ endif
 .PHONY: libedgetpu \
         libedgetpu-direct \
         libedgetpu-throttled \
-        workspace \
+        deb \
+        deb-armhf \
+        deb-arm64 \
         clean
 
 libedgetpu: libedgetpu-direct libedgetpu-throttled
 
-libedgetpu-direct: workspace
+libedgetpu-direct:
 	bazel build $(BAZEL_BUILD_FLAGS) $(BAZEL_BUILD_TARGET)
 	$(call copy_out,direct)
 	$(call strip_out,direct)
 
-libedgetpu-throttled: workspace
+libedgetpu-throttled:
 	bazel build $(BAZEL_BUILD_FLAGS) --copt=-DTHROTTLE_EDGE_TPU $(BAZEL_BUILD_TARGET)
 	$(call copy_out,throttled)
 	$(call strip_out,throttled)
 
-workspace: bazel/WORKSPACE bazel/$(WORKSPACE_PLATFORM_FILE)
-	cat $^ > WORKSPACE
+deb:
+	dpkg-buildpackage -rfakeroot -us -uc -tc -b
+
+deb-armhf:
+	dpkg-buildpackage -rfakeroot -us -uc -tc -b -a armhf -d
+
+deb-arm64:
+	dpkg-buildpackage -rfakeroot -us -uc -tc -b -a arm64 -d
 
 clean:
 	rm -rf $(OUT_DIR)

WORKSPACE

@@ -0,0 +1,23 @@
+# Copyright 2019 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.
+workspace(name = "libedgetpu")
+
+load(":workspace.bzl", "libedgetpu_dependencies")
+
+libedgetpu_dependencies()
+load("@org_tensorflow//tensorflow:workspace.bzl", "tf_workspace")
+tf_workspace(tf_repo_name = "org_tensorflow")
+
+load("@coral_crosstool//:configure.bzl", "cc_crosstool")
+cc_crosstool(name = "crosstool")

api/BUILD

@@ -12,6 +12,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+load("//:libedgetpu_cc_rules.bzl", "libedgetpu_cc_library")
+
 # Description:
 #   Darwinn API headers
 load(
@@ -23,12 +25,12 @@ package(default_visibility = ["//visibility:public"])
 
 licenses(["notice"])
 
-cc_library(
+libedgetpu_cc_library(
     name = "chip",
     hdrs = ["chip.h"],
 )
 
-cc_library(
+libedgetpu_cc_library(
     name = "tensor_util",
     srcs = ["tensor_util.cc"],
     hdrs = ["tensor_util.h"],
@@ -39,7 +41,7 @@ cc_library(
     ],
 )
 
-cc_library(
+libedgetpu_cc_library(
     name = "layer_information",
     srcs = ["layer_information.cc"],
     hdrs = ["layer_information.h"],
@@ -51,7 +53,7 @@ cc_library(
     ],
 )
 
-cc_library(
+libedgetpu_cc_library(
     name = "request",
     hdrs = ["request.h"],
     deps = [
@@ -60,7 +62,7 @@ cc_library(
     ],
 )
 
-cc_library(
+libedgetpu_cc_library(
     name = "driver",
     hdrs = ["driver.h"],
     deps = [
@@ -73,7 +75,7 @@ cc_library(
     ],
 )
 
-cc_library(
+libedgetpu_cc_library(
     name = "driver_options_helper",
     srcs = ["driver_options_helper.cc"],
     hdrs = ["driver_options_helper.h"],
@@ -83,20 +85,20 @@ cc_library(
     ],
 )
 
-cc_library(
+libedgetpu_cc_library(
     name = "allocated_buffer",
     srcs = ["allocated_buffer.cc"],
     hdrs = ["allocated_buffer.h"],
     deps = ["//port"],
 )
 
-cc_library(
+libedgetpu_cc_library(
     name = "dram_buffer",
     hdrs = ["dram_buffer.h"],
     deps = ["//port"],
 )
 
-cc_library(
+libedgetpu_cc_library(
     name = "buffer",
     srcs = ["buffer.cc"],
     hdrs = ["buffer.h"],
@@ -107,7 +109,7 @@ cc_library(
     ],
 )
 
-cc_library(
+libedgetpu_cc_library(
     name = "driver_factory",
     srcs = ["driver_factory.cc"],
     hdrs = ["driver_factory.h"],
@@ -119,7 +121,7 @@ cc_library(
     ],
 )
 
-cc_library(
+libedgetpu_cc_library(
     name = "package_reference",
     hdrs = ["package_reference.h"],
     deps = [
@@ -130,7 +132,7 @@ cc_library(
     ],
 )
 
-cc_library(
+libedgetpu_cc_library(
     name = "runtime_version",
     hdrs = ["runtime_version.h"],
 )
@@ -141,15 +143,15 @@ flatbuffer_cc_library(
     flatc_args = [""],
 )
 
-cc_library(
+libedgetpu_cc_library(
     name = "timing",
     hdrs = ["timing.h"],
     deps = [
         "//port",
     ],
 )
 
-cc_library(
+libedgetpu_cc_library(
     name = "watchdog",
     srcs = ["watchdog.cc"],
     hdrs = ["watchdog.h"],
@@ -161,7 +163,7 @@ cc_library(
     ],
 )
 
-cc_library(
+libedgetpu_cc_library(
     name = "telemeter_interface",
     hdrs = [
         "telemeter_interface.h",
@@ -171,7 +173,7 @@ cc_library(
     ],
 )
 
-cc_library(
+libedgetpu_cc_library(
     name = "execution_context_interface",
     hdrs = [
         "execution_context_interface.h",

api/driver.h

@@ -85,8 +85,7 @@ class Driver {
   // integer values may be different from those in NNAPI or other APIs. The
   // values here are defined in the order of priority when there are multiple
   // models requesting different preferences (e.g. sustained speed takes
-  // priority over low power). For more information, please see:
-  // http://go/noronha-execution-preference
+  // priority over low power).
   enum class ExecutionPreference {
     // Run at the absolute maximum performance.
     kSingleFastAnswer = 0,

api/layer_information.cc

@@ -203,30 +203,41 @@ bool OutputLayerInformation::NeedsRelayout() const {
 // TODO Add unit tests for this method.
 util::Status OutputLayerInformation::Relayout(unsigned char* dest,
                                               const unsigned char* src) const {
-  // TODO: re-use the same buffer and avoid an unnecessary memcopy
-  // when relayout is not needed.
-  if (!NeedsRelayout()) {
-    memcpy(dest, src,
-           batch_dim() * y_dim() * x_dim() * z_dim() * DataTypeSize());
-    return util::OkStatus();
-  }
-
-  if (output_layer_->shape_info()) {
-    // If output shape info exists in the executable, use the new re-layout
-    // function. Currently, this is only enabled for models with multiple
-    // batches.
-    return RelayoutWithShapeInformation(dest, src);
-  }
-
   const auto data_type_size = DataTypeSize();
   const int z_bytes = z_dim() * data_type_size;
+  const int executions = execution_count_per_inference();
+
+  if (executions == 1) {
+    // Handle case when execution count is equal to 1, since if execution count
+    // is greater than 1, there might be padding data in-between.
+
+    // TODO: re-use the same buffer and avoid an unnecessary
+    // memcopy when relayout is not needed.
+    if (!NeedsRelayout()) {
+      memcpy(dest, src, batch_dim() * y_dim() * x_dim() * z_bytes);
+      return util::OkStatus();
+    }
+
+    if (output_layer_->shape_info()) {
+      // If output shape info exists in the executable, use the new re-layout
+      // function. Currently, this is only enabled for models with multiple
+      // batches.
+      return RelayoutWithShapeInformation(dest, src);
+    }
+  } else if (PaddedSizeBytes() == ActualSizeBytes() && !NeedsRelayout()) {
+    // Use memcpy if `executions` is greater than 1 and there is no internal
+    // padding between iterations.
+    if (dest != src) {
+      memcpy(dest, src, ActualSizeBytes());
+    }
+    return util::OkStatus();
+  }
 
   if (y_dim() == 1 && x_dim() == 1) {
     // One dimensional output (only z-dimension).
     if (src != dest) {
       const int padded_size_bytes = PaddedSizeBytes();
       const int actual_size_bytes = ActualSizeBytes();
-      const int executions = execution_count_per_inference();
       if (executions == 1 || padded_size_bytes == actual_size_bytes) {
         memcpy(dest, src, z_bytes * executions);
       } else {
@@ -276,8 +287,7 @@ util::Status OutputLayerInformation::Relayout(unsigned char* dest,
 // provided we have a guaranteed way of ensuring this function would be inlined
 // so that the compiler optimizations based on compile-time-constants can kick
 // in.
-#define RELAYOUT_WITH_Z_BYTES_SPECIALIZATION(                                 \
-    num_z_bytes, num_z_bytes_padded)                                          \
+#define RELAYOUT_WITH_Z_BYTES_SPECIALIZATION(num_z_bytes, num_z_bytes_padded) \
   do {                                                                        \
     for (int y = 0; y < y_dim(); ++y) {                                       \
       const auto y_buffer_index = GetYBufferIndex(y);                         \
@@ -330,6 +340,12 @@ util::Status OutputLayerInformation::Relayout(unsigned char* dest,
           active_tile_x_sizes.size() > 1 || first_y_tile != last_y_tile;
 
       if (need_relayout) {
+        // TODO: If iteration count is more than 1, we need to make
+        // sure we advance 'src' and 'dest' correctly due to padding issue. We
+        // don't have test case now.
+        CHECK_EQ(executions, 1)
+            << "Verification is missing if execution count is greater than 1";
+
         // If there's no z padding, copy one xz block on one tile at a time.
         for (int y = 0; y < y_dim(); ++y) {
           const auto y_buffer_index = GetYBufferIndex(y);
@@ -347,17 +363,11 @@ util::Status OutputLayerInformation::Relayout(unsigned char* dest,
         }
       } else {
         // TODO: avoid copy and assign in caller directly.
-        memcpy(dest, src, x_dim() * y_dim() * z_bytes);
+        memcpy(dest, src, x_dim() * y_dim() * z_bytes * executions);
       }
     }
 
 #undef RELAYOUT_WITH_Z_BYTES_SPECIALIZATION
-
-    // TODO: If iteration count is more than 1, we need to make sure we
-    // advance 'src' and 'dest' correctly due to padding issue. We don't have
-    // test case now.
-    CHECK_EQ(execution_count_per_inference(), 1)
-        << "Verification is missing if execution count is greater than 1";
   }
 
   return util::OkStatus();