Fuse resize and mirror padding ops into convolutions

Spatial transformations like padding and bilinear resizing can be merged into the im2col stage of conv2d. This reduces the memory usage considerably (from 338MB to 224MB) and latency (by 15%) on some models, and helps us avoid OOM crashes on iOS. This PR has all the changes needed to fuse these particular ops, including the kernels themselves and integration into the optimize_for_inference script.
Change: 132094335

1 files changed, 240 insertions, 0 deletions

diff --git a/tensorflow/core/kernels/conv_ops_test.cc b/tensorflow/core/kernels/conv_ops_test.cc
new file mode 100644
index 0000000000..228f2d5def
--- /dev/null
+++ b/tensorflow/core/kernels/conv_ops_test.cc
@@ -0,0 +1,240 @@
+/* Copyright 2015 The TensorFlow Authors. All Rights Reserved.
+
+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
+
+    http://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.
+==============================================================================*/
+
+#include "tensorflow/cc/ops/const_op.h"
+#include "tensorflow/cc/ops/image_ops.h"
+#include "tensorflow/cc/ops/nn_ops.h"
+#include "tensorflow/cc/ops/standard_ops.h"
+#include "tensorflow/core/common_runtime/kernel_benchmark_testlib.h"
+#include "tensorflow/core/framework/fake_input.h"
+#include "tensorflow/core/framework/node_def_builder.h"
+#include "tensorflow/core/framework/tensor.h"
+#include "tensorflow/core/framework/types.pb.h"
+#include "tensorflow/core/kernels/ops_testutil.h"
+#include "tensorflow/core/kernels/ops_util.h"
+#include "tensorflow/core/platform/test.h"
+#include "tensorflow/core/platform/test_benchmark.h"
+#include "tensorflow/core/public/session.h"
+
+namespace tensorflow {
+
+class FusedResizePadConvOpTest : public OpsTestBase {
+ protected:
+  void HandwrittenConv() {
+    const int stride = 1;
+    TF_EXPECT_OK(NodeDefBuilder("fused_resize_op", "FusedResizeAndPadConv2D")
+                     .Input(FakeInput(DT_FLOAT))
+                     .Input(FakeInput(DT_INT32))
+                     .Input(FakeInput(DT_INT32))
+                     .Input(FakeInput(DT_FLOAT))
+                     .Attr("T", DT_FLOAT)
+                     .Attr("resize_align_corners", false)
+                     .Attr("mode", "REFLECT")
+                     .Attr("strides", {1, stride, stride, 1})
+                     .Attr("padding", "SAME")
+                     .Finalize(node_def()));
+    TF_EXPECT_OK(InitOp());
+    const int depth = 1;
+    const int image_width = 4;
+    const int image_height = 3;
+    const int image_batch_count = 1;
+    // The image matrix is:
+    // |  1 |  2 |  3 |  4 |
+    // |  5 |  6 |  7 |  8 |
+    // |  9 | 10 | 11 | 12 |
+    Tensor image(DT_FLOAT,
+                 {image_batch_count, image_height, image_width, depth});
+    test::FillValues<float>(&image, {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12});
+
+    // The filter matrix is:
+    // | 1 | 4 | 7 |
+    // | 2 | 5 | 8 |
+    // | 3 | 6 | 9 |
+    const int filter_size = 3;
+    const int filter_count = 1;
+    Tensor filter(DT_FLOAT, {filter_size, filter_size, depth, filter_count});
+    test::FillValues<float>(&filter, {1, 4, 7, 2, 5, 8, 3, 6, 9});
+
+    const int resized_width = image_width;
+    const int resized_height = image_height;
+
+    const int top_padding = 0;
+    const int bottom_padding = 0;
+    const int left_padding = 0;
+    const int right_padding = 0;
+
+    AddInputFromArray<float>(image.shape(), image.flat<float>());
+    AddInputFromArray<int32>(TensorShape({2}), {resized_height, resized_width});
+    AddInputFromArray<int32>(
+        TensorShape({4, 2}),
+        {0, 0, top_padding, bottom_padding, left_padding, right_padding, 0, 0});
+    AddInputFromArray<float>(filter.shape(), filter.flat<float>());
+    TF_ASSERT_OK(RunOpKernel());
+
+    // We're sliding the 3x3 filter across the 3x4 image, with accesses outside
+    // the input set to zero because we're using the 'SAME' padding mode.
+    // The calculations behind the expected output are:
+    // (1*0)+(4*0)+(7*0)+(2*0)+(5*1)+(8*2)+(3*0)+(6*5)+(9*6)=105
+    // (1*0)+(4*0)+(7*0)+(2*1)+(5*2)+(8*3)+(3*5)+(6*6)+(9*7)=150
+    // (1*0)+(4*0)+(7*0)+(2*2)+(5*3)+(8*4)+(3*6)+(6*7)+(9*8)=183
+    // (1*0)+(4*0)+(7*0)+(2*3)+(5*4)+(8*0)+(3*7)+(6*8)+(9*0)=95
+    // (1*0)+(4*1)+(7*2)+(2*0)+(5*5)+(8*6)+(3*0)+(6*9)+(9*10)=235
+    // (1*1)+(4*2)+(7*3)+(2*5)+(5*6)+(8*7)+(3*9)+(6*10)+(9*11)=312
+    // (1*2)+(4*3)+(7*4)+(2*6)+(5*7)+(8*8)+(3*10)+(6*11)+(9*12)=357
+    // (1*3)+(4*4)+(7*0)+(2*7)+(5*8)+(8*0)+(3*11)+(6*12)+(9*0)=178
+    // (1*0)+(4*5)+(7*6)+(2*0)+(5*9)+(8*10)+(3*0)+(6*0)+(9*0)=187
+    // (1*5)+(4*6)+(7*7)+(2*9)+(5*10)+(8*11)+(3*0)+(6*0)+(9*0)=234
+    // (1*6)+(4*7)+(7*8)+(2*10)+(5*11)+(8*12)+(3*0)+(6*0)+(9*0)=261
+    // (1*7)+(4*11)+(7*0)+(2*8)+(5*12)+(8*0)+(3*0)+(6*0)+(9*0)=121
+    // This means we should end up with this matrix:
+    // |  105  |  150  |  183  |   95  |
+    // |  235  |  312  |  357  |  178  |
+    // |  187  |  234  |  261  |  121  |
+    const int expected_width = image_width;
+    const int expected_height = image_height * filter_count;
+    Tensor expected(DT_FLOAT, TensorShape({image_batch_count, expected_height,
+                                           expected_width, filter_count}));
+    test::FillValues<float>(
+        &expected, {105, 150, 183, 95, 235, 312, 357, 178, 187, 234, 261, 121});
+    const Tensor& output = *GetOutput(0);
+    test::ExpectTensorNear<float>(expected, output, 1e-5);
+  }
+
+  void CompareFusedAndSeparate(int input_width, int input_height,
+                               int input_depth, int resize_width,
+                               int resize_height, int y_padding, int x_padding,
+                               int filter_size, int filter_count,
+                               bool resize_align_corners, string pad_mode,
+                               int stride, string padding) {
+    auto root = tensorflow::Scope::NewRootScope();
+    using namespace ::tensorflow::ops;  // NOLINT(build/namespaces)
+
+    const size_t input_data_size = input_height * input_width * input_depth;
+    Tensor input_data(DT_FLOAT,
+                      TensorShape({1, input_height, input_width, input_depth}));
+    for (int i = 0; i < input_data_size; ++i) {
+      input_data.flat<float>()(i) = i + 1.0f;
+    }
+    Output input =
+        Const(root.WithOpName("input"), Input::Initializer(input_data));
+
+    const size_t filter_data_size =
+        filter_size * filter_size * filter_count * input_depth;
+    Tensor filter_data(DT_FLOAT, TensorShape({filter_size, filter_size,
+                                              input_depth, filter_count}));
+    for (int i = 0; i < filter_data_size; ++i) {
+      filter_data.flat<float>()(i) = i + 1.0f;
+    }
+    Output filter =
+        Const(root.WithOpName("filter"), Input::Initializer(filter_data));
+
+    Output resize_size =
+        Const(root.WithOpName("resize_size"), {resize_height, resize_width});
+    Output resize =
+        ResizeBilinear(root.WithOpName("resize"), input, resize_size,
+                       ResizeBilinear::AlignCorners(resize_align_corners));
+    Output paddings =
+        Const(root.WithOpName("paddings"),
+              {{0, 0}, {y_padding, y_padding}, {x_padding, x_padding}, {0, 0}});
+    Output mirror_pad =
+        MirrorPad(root.WithOpName("mirror_pad"), resize, paddings, pad_mode);
+    Output conv = Conv2D(root.WithOpName("conv"), mirror_pad, filter,
+                         {1, stride, stride, 1}, padding);
+
+    Output fused_conv = FusedResizeAndPadConv2D(
+        root.WithOpName("fused_conv"), input, resize_size, paddings, filter,
+        pad_mode, {1, stride, stride, 1}, padding,
+        FusedResizeAndPadConv2D::ResizeAlignCorners(resize_align_corners));
+
+    tensorflow::GraphDef graph;
+    TF_ASSERT_OK(root.ToGraphDef(&graph));
+
+    std::unique_ptr<tensorflow::Session> session(
+        tensorflow::NewSession(tensorflow::SessionOptions()));
+    TF_ASSERT_OK(session->Create(graph));
+
+    std::vector<Tensor> unfused_tensors;
+    TF_ASSERT_OK(session->Run({}, {"conv"}, {}, &unfused_tensors));
+
+    std::vector<Tensor> fused_tensors;
+    TF_ASSERT_OK(session->Run({}, {"fused_conv"}, {}, &fused_tensors));
+
+    test::ExpectTensorNear<float>(unfused_tensors[0], fused_tensors[0], 1e-5);
+  }
+};
+
+TEST_F(FusedResizePadConvOpTest, HandwrittenConv) { HandwrittenConv(); }
+
+TEST_F(FusedResizePadConvOpTest, IdentityComparative) {
+  CompareFusedAndSeparate(10, 10, 1, 10, 10, 0, 0, 1, 1, false, "REFLECT", 1,
+                          "SAME");
+}
+
+TEST_F(FusedResizePadConvOpTest, ConvOnlyComparative) {
+  CompareFusedAndSeparate(10, 10, 3, 10, 10, 0, 0, 4, 4, false, "REFLECT", 1,
+                          "SAME");
+}
+
+TEST_F(FusedResizePadConvOpTest, ResizeOnlyComparative) {
+  CompareFusedAndSeparate(10, 10, 1, 20, 20, 0, 0, 1, 1, false, "REFLECT", 1,
+                          "SAME");
+}
+
+TEST_F(FusedResizePadConvOpTest, ResizeAndConvComparative) {
+  CompareFusedAndSeparate(2, 2, 4, 4, 2, 0, 0, 2, 2, false, "REFLECT", 1,
+                          "SAME");
+}
+
+TEST_F(FusedResizePadConvOpTest, ResizeAlignAndConvComparative) {
+  CompareFusedAndSeparate(2, 2, 4, 4, 2, 0, 0, 2, 2, true, "REFLECT", 1,
+                          "SAME");
+}
+
+TEST_F(FusedResizePadConvOpTest, ResizeAndConvStridedComparative) {
+  CompareFusedAndSeparate(2, 2, 4, 4, 2, 0, 0, 2, 2, false, "REFLECT", 2,
+                          "SAME");
+}
+
+TEST_F(FusedResizePadConvOpTest, ResizeAlignAndConvValidComparative) {
+  CompareFusedAndSeparate(2, 2, 4, 4, 2, 0, 0, 2, 2, true, "REFLECT", 1,
+                          "VALID");
+}
+
+TEST_F(FusedResizePadConvOpTest, PadOnlyComparative) {
+  CompareFusedAndSeparate(4, 4, 1, 4, 4, 2, 2, 1, 1, false, "REFLECT", 1,
+                          "SAME");
+}
+
+TEST_F(FusedResizePadConvOpTest, PadOnlyWithChannelsComparative) {
+  CompareFusedAndSeparate(4, 4, 3, 4, 4, 2, 2, 1, 1, false, "REFLECT", 1,
+                          "SAME");
+}
+
+TEST_F(FusedResizePadConvOpTest, ResizeAndPadComparative) {
+  CompareFusedAndSeparate(4, 4, 1, 6, 6, 2, 2, 1, 1, false, "REFLECT", 1,
+                          "SAME");
+}
+
+TEST_F(FusedResizePadConvOpTest, PadOnlySymmetricComparative) {
+  CompareFusedAndSeparate(4, 4, 1, 4, 4, 2, 2, 1, 1, false, "SYMMETRIC", 1,
+                          "SAME");
+}
+
+TEST_F(FusedResizePadConvOpTest, ResizeAndPadSymmetricComparative) {
+  CompareFusedAndSeparate(4, 4, 3, 6, 6, 2, 2, 1, 1, false, "SYMMETRIC", 1,
+                          "SAME");
+}
+
+}  // namespace tensorflow