Project import generated by Copybara.

GitOrigin-RevId: d0039a576e2db9c0fcefffd26a527df74cbe145b

mediapipe/calculators/image/image_properties_calculator.cc

@@ -19,6 +19,11 @@
 #include "mediapipe/gpu/gpu_buffer.h"
 #endif  //  !MEDIAPIPE_DISABLE_GPU
 
+namespace {
+constexpr char kImageFrameTag[] = "IMAGE";
+constexpr char kGpuBufferTag[] = "IMAGE_GPU";
+}  // namespace
+
 namespace mediapipe {
 
 // Extracts image properties from the input image and outputs the properties.
@@ -40,13 +45,14 @@ namespace mediapipe {
 class ImagePropertiesCalculator : public CalculatorBase {
  public:
   static ::mediapipe::Status GetContract(CalculatorContract* cc) {
-    RET_CHECK(cc->Inputs().HasTag("IMAGE") ^ cc->Inputs().HasTag("IMAGE_GPU"));
-    if (cc->Inputs().HasTag("IMAGE")) {
-      cc->Inputs().Tag("IMAGE").Set<ImageFrame>();
+    RET_CHECK(cc->Inputs().HasTag(kImageFrameTag) ^
+              cc->Inputs().HasTag(kGpuBufferTag));
+    if (cc->Inputs().HasTag(kImageFrameTag)) {
+      cc->Inputs().Tag(kImageFrameTag).Set<ImageFrame>();
     }
 #if !defined(MEDIAPIPE_DISABLE_GPU)
-    if (cc->Inputs().HasTag("IMAGE_GPU")) {
-      cc->Inputs().Tag("IMAGE_GPU").Set<::mediapipe::GpuBuffer>();
+    if (cc->Inputs().HasTag(kGpuBufferTag)) {
+      cc->Inputs().Tag(kGpuBufferTag).Set<::mediapipe::GpuBuffer>();
     }
 #endif  //  !MEDIAPIPE_DISABLE_GPU
 
@@ -66,16 +72,17 @@ class ImagePropertiesCalculator : public CalculatorBase {
     int width;
     int height;
 
-    if (cc->Inputs().HasTag("IMAGE") && !cc->Inputs().Tag("IMAGE").IsEmpty()) {
-      const auto& image = cc->Inputs().Tag("IMAGE").Get<ImageFrame>();
+    if (cc->Inputs().HasTag(kImageFrameTag) &&
+        !cc->Inputs().Tag(kImageFrameTag).IsEmpty()) {
+      const auto& image = cc->Inputs().Tag(kImageFrameTag).Get<ImageFrame>();
       width = image.Width();
       height = image.Height();
     }
 #if !defined(MEDIAPIPE_DISABLE_GPU)
-    if (cc->Inputs().HasTag("IMAGE_GPU") &&
-        !cc->Inputs().Tag("IMAGE_GPU").IsEmpty()) {
+    if (cc->Inputs().HasTag(kGpuBufferTag) &&
+        !cc->Inputs().Tag(kGpuBufferTag).IsEmpty()) {
       const auto& image =
-          cc->Inputs().Tag("IMAGE_GPU").Get<mediapipe::GpuBuffer>();
+          cc->Inputs().Tag(kGpuBufferTag).Get<mediapipe::GpuBuffer>();
       width = image.width();
       height = image.height();
     }

mediapipe/calculators/image/image_transformation_calculator.cc

@@ -47,6 +47,9 @@ namespace mediapipe {
 #endif  //  !MEDIAPIPE_DISABLE_GPU
 
 namespace {
+constexpr char kImageFrameTag[] = "IMAGE";
+constexpr char kGpuBufferTag[] = "IMAGE_GPU";
+
 int RotationModeToDegrees(mediapipe::RotationMode_Mode rotation) {
   switch (rotation) {
     case mediapipe::RotationMode_Mode_UNKNOWN:
@@ -95,7 +98,7 @@ mediapipe::ScaleMode_Mode ParseScaleMode(
 // Scales, rotates, and flips images horizontally or vertically.
 //
 // Input:
-//   One of the following two tags:
+//   One of the following tags:
 //   IMAGE: ImageFrame representing the input image.
 //   IMAGE_GPU: GpuBuffer representing the input image.
 //
@@ -113,7 +116,7 @@ mediapipe::ScaleMode_Mode ParseScaleMode(
 //   corresponding field in the calculator options.
 //
 // Output:
-//   One of the following two tags:
+//   One of the following tags:
 //   IMAGE - ImageFrame representing the output image.
 //   IMAGE_GPU - GpuBuffer representing the output image.
 //
@@ -152,7 +155,8 @@ mediapipe::ScaleMode_Mode ParseScaleMode(
 // Note: To enable horizontal or vertical flipping, specify them in the
 // calculator options. Flipping is applied after rotation.
 //
-// Note: Only scale mode STRETCH is currently supported on CPU.
+// Note: Input defines output, so only matchig types supported:
+// IMAGE -> IMAGE  or  IMAGE_GPU -> IMAGE_GPU
 //
 class ImageTransformationCalculator : public CalculatorBase {
  public:
@@ -186,7 +190,7 @@ class ImageTransformationCalculator : public CalculatorBase {
 
   bool use_gpu_ = false;
 #if !defined(MEDIAPIPE_DISABLE_GPU)
-  GlCalculatorHelper helper_;
+  GlCalculatorHelper gpu_helper_;
   std::unique_ptr<QuadRenderer> rgb_renderer_;
   std::unique_ptr<QuadRenderer> yuv_renderer_;
   std::unique_ptr<QuadRenderer> ext_rgb_renderer_;
@@ -197,21 +201,22 @@ REGISTER_CALCULATOR(ImageTransformationCalculator);
 // static
 ::mediapipe::Status ImageTransformationCalculator::GetContract(
     CalculatorContract* cc) {
-  RET_CHECK(cc->Inputs().HasTag("IMAGE") ^ cc->Inputs().HasTag("IMAGE_GPU"));
-  RET_CHECK(cc->Outputs().HasTag("IMAGE") ^ cc->Outputs().HasTag("IMAGE_GPU"));
+  // Only one input can be set, and the output type must match.
+  RET_CHECK(cc->Inputs().HasTag(kImageFrameTag) ^
+            cc->Inputs().HasTag(kGpuBufferTag));
 
   bool use_gpu = false;
 
-  if (cc->Inputs().HasTag("IMAGE")) {
-    RET_CHECK(cc->Outputs().HasTag("IMAGE"));
-    cc->Inputs().Tag("IMAGE").Set<ImageFrame>();
-    cc->Outputs().Tag("IMAGE").Set<ImageFrame>();
+  if (cc->Inputs().HasTag(kImageFrameTag)) {
+    RET_CHECK(cc->Outputs().HasTag(kImageFrameTag));
+    cc->Inputs().Tag(kImageFrameTag).Set<ImageFrame>();
+    cc->Outputs().Tag(kImageFrameTag).Set<ImageFrame>();
   }
 #if !defined(MEDIAPIPE_DISABLE_GPU)
-  if (cc->Inputs().HasTag("IMAGE_GPU")) {
-    RET_CHECK(cc->Outputs().HasTag("IMAGE_GPU"));
-    cc->Inputs().Tag("IMAGE_GPU").Set<GpuBuffer>();
-    cc->Outputs().Tag("IMAGE_GPU").Set<GpuBuffer>();
+  if (cc->Inputs().HasTag(kGpuBufferTag)) {
+    RET_CHECK(cc->Outputs().HasTag(kGpuBufferTag));
+    cc->Inputs().Tag(kGpuBufferTag).Set<GpuBuffer>();
+    cc->Outputs().Tag(kGpuBufferTag).Set<GpuBuffer>();
     use_gpu |= true;
   }
 #endif  //  !MEDIAPIPE_DISABLE_GPU
@@ -259,7 +264,7 @@ REGISTER_CALCULATOR(ImageTransformationCalculator);
 
   options_ = cc->Options<ImageTransformationCalculatorOptions>();
 
-  if (cc->Inputs().HasTag("IMAGE_GPU")) {
+  if (cc->Inputs().HasTag(kGpuBufferTag)) {
     use_gpu_ = true;
   }
 
@@ -300,7 +305,7 @@ REGISTER_CALCULATOR(ImageTransformationCalculator);
   if (use_gpu_) {
 #if !defined(MEDIAPIPE_DISABLE_GPU)
     // Let the helper access the GL context information.
-    MP_RETURN_IF_ERROR(helper_.Open(cc));
+    MP_RETURN_IF_ERROR(gpu_helper_.Open(cc));
 #else
     RET_CHECK_FAIL() << "GPU processing not enabled.";
 #endif  //  !MEDIAPIPE_DISABLE_GPU
@@ -328,18 +333,14 @@ REGISTER_CALCULATOR(ImageTransformationCalculator);
 
   if (use_gpu_) {
 #if !defined(MEDIAPIPE_DISABLE_GPU)
-    if (cc->Inputs().Tag("IMAGE_GPU").IsEmpty()) {
-      // Image is missing, hence no way to produce output image. (Timestamp
-      // bound will be updated automatically.)
+    if (cc->Inputs().Tag(kGpuBufferTag).IsEmpty()) {
       return ::mediapipe::OkStatus();
     }
-    return helper_.RunInGlContext(
+    return gpu_helper_.RunInGlContext(
         [this, cc]() -> ::mediapipe::Status { return RenderGpu(cc); });
 #endif  //  !MEDIAPIPE_DISABLE_GPU
   } else {
-    if (cc->Inputs().Tag("IMAGE").IsEmpty()) {
-      // Image is missing, hence no way to produce output image. (Timestamp
-      // bound will be updated automatically.)
+    if (cc->Inputs().Tag(kImageFrameTag).IsEmpty()) {
       return ::mediapipe::OkStatus();
     }
     return RenderCpu(cc);
@@ -354,7 +355,7 @@ REGISTER_CALCULATOR(ImageTransformationCalculator);
     QuadRenderer* rgb_renderer = rgb_renderer_.release();
     QuadRenderer* yuv_renderer = yuv_renderer_.release();
     QuadRenderer* ext_rgb_renderer = ext_rgb_renderer_.release();
-    helper_.RunInGlContext([rgb_renderer, yuv_renderer, ext_rgb_renderer] {
+    gpu_helper_.RunInGlContext([rgb_renderer, yuv_renderer, ext_rgb_renderer] {
       if (rgb_renderer) {
         rgb_renderer->GlTeardown();
         delete rgb_renderer;
@@ -376,17 +377,21 @@ REGISTER_CALCULATOR(ImageTransformationCalculator);
 
 ::mediapipe::Status ImageTransformationCalculator::RenderCpu(
     CalculatorContext* cc) {
-  const auto& input_img = cc->Inputs().Tag("IMAGE").Get<ImageFrame>();
-  cv::Mat input_mat = formats::MatView(&input_img);
-  cv::Mat scaled_mat;
+  cv::Mat input_mat;
+  mediapipe::ImageFormat::Format format;
 
-  const int input_width = input_img.Width();
-  const int input_height = input_img.Height();
+  const auto& input = cc->Inputs().Tag(kImageFrameTag).Get<ImageFrame>();
+  input_mat = formats::MatView(&input);
+  format = input.Format();
+
+  const int input_width = input_mat.cols;
+  const int input_height = input_mat.rows;
   if (!output_height_ || !output_width_) {
     output_height_ = input_height;
     output_width_ = input_width;
   }
 
+  cv::Mat scaled_mat;
   if (scale_mode_ == mediapipe::ScaleMode_Mode_STRETCH) {
     cv::resize(input_mat, scaled_mat, cv::Size(output_width_, output_height_));
   } else {
@@ -443,10 +448,12 @@ REGISTER_CALCULATOR(ImageTransformationCalculator);
   }
 
   std::unique_ptr<ImageFrame> output_frame(
-      new ImageFrame(input_img.Format(), output_width, output_height));
+      new ImageFrame(format, output_width, output_height));
   cv::Mat output_mat = formats::MatView(output_frame.get());
   flipped_mat.copyTo(output_mat);
-  cc->Outputs().Tag("IMAGE").Add(output_frame.release(), cc->InputTimestamp());
+  cc->Outputs()
+      .Tag(kImageFrameTag)
+      .Add(output_frame.release(), cc->InputTimestamp());
 
   return ::mediapipe::OkStatus();
 }
@@ -454,7 +461,7 @@ REGISTER_CALCULATOR(ImageTransformationCalculator);
 ::mediapipe::Status ImageTransformationCalculator::RenderGpu(
     CalculatorContext* cc) {
 #if !defined(MEDIAPIPE_DISABLE_GPU)
-  const auto& input = cc->Inputs().Tag("IMAGE_GPU").Get<GpuBuffer>();
+  const auto& input = cc->Inputs().Tag(kGpuBufferTag).Get<GpuBuffer>();
   const int input_width = input.width();
   const int input_height = input.height();
 
@@ -485,11 +492,11 @@ REGISTER_CALCULATOR(ImageTransformationCalculator);
                                  {"video_frame_y", "video_frame_uv"}));
     }
     renderer = yuv_renderer_.get();
-    src1 = helper_.CreateSourceTexture(input, 0);
+    src1 = gpu_helper_.CreateSourceTexture(input, 0);
   } else  // NOLINT(readability/braces)
 #endif    // iOS
   {
-    src1 = helper_.CreateSourceTexture(input);
+    src1 = gpu_helper_.CreateSourceTexture(input);
 #if defined(TEXTURE_EXTERNAL_OES)
     if (src1.target() == GL_TEXTURE_EXTERNAL_OES) {
       if (!ext_rgb_renderer_) {
@@ -515,10 +522,10 @@ REGISTER_CALCULATOR(ImageTransformationCalculator);
   mediapipe::FrameRotation rotation =
       mediapipe::FrameRotationFromDegrees(RotationModeToDegrees(rotation_));
 
-  auto dst = helper_.CreateDestinationTexture(output_width, output_height,
-                                              input.format());
+  auto dst = gpu_helper_.CreateDestinationTexture(output_width, output_height,
+                                                  input.format());
 
-  helper_.BindFramebuffer(dst);  // GL_TEXTURE0
+  gpu_helper_.BindFramebuffer(dst);  // GL_TEXTURE0
   glActiveTexture(GL_TEXTURE1);
   glBindTexture(src1.target(), src1.name());
 
@@ -533,8 +540,8 @@ REGISTER_CALCULATOR(ImageTransformationCalculator);
   // Execute GL commands, before getting result.
   glFlush();
 
-  auto output = dst.GetFrame<GpuBuffer>();
-  cc->Outputs().Tag("IMAGE_GPU").Add(output.release(), cc->InputTimestamp());
+  auto output = dst.template GetFrame<GpuBuffer>();
+  cc->Outputs().Tag(kGpuBufferTag).Add(output.release(), cc->InputTimestamp());
 
 #endif  //  !MEDIAPIPE_DISABLE_GPU
 

mediapipe/calculators/image/recolor_calculator.cc

@@ -32,6 +32,11 @@
 
 namespace {
 enum { ATTRIB_VERTEX, ATTRIB_TEXTURE_POSITION, NUM_ATTRIBUTES };
+
+constexpr char kImageFrameTag[] = "IMAGE";
+constexpr char kMaskCpuTag[] = "MASK";
+constexpr char kGpuBufferTag[] = "IMAGE_GPU";
+constexpr char kMaskGpuTag[] = "MASK_GPU";
 }  // namespace
 
 namespace mediapipe {
@@ -112,39 +117,41 @@ REGISTER_CALCULATOR(RecolorCalculator);
   bool use_gpu = false;
 
 #if !defined(MEDIAPIPE_DISABLE_GPU)
-  if (cc->Inputs().HasTag("IMAGE_GPU")) {
-    cc->Inputs().Tag("IMAGE_GPU").Set<mediapipe::GpuBuffer>();
+  if (cc->Inputs().HasTag(kGpuBufferTag)) {
+    cc->Inputs().Tag(kGpuBufferTag).Set<mediapipe::GpuBuffer>();
     use_gpu |= true;
   }
 #endif  //  !MEDIAPIPE_DISABLE_GPU
-  if (cc->Inputs().HasTag("IMAGE")) {
-    cc->Inputs().Tag("IMAGE").Set<ImageFrame>();
+  if (cc->Inputs().HasTag(kImageFrameTag)) {
+    cc->Inputs().Tag(kImageFrameTag).Set<ImageFrame>();
   }
 
 #if !defined(MEDIAPIPE_DISABLE_GPU)
-  if (cc->Inputs().HasTag("MASK_GPU")) {
-    cc->Inputs().Tag("MASK_GPU").Set<mediapipe::GpuBuffer>();
+  if (cc->Inputs().HasTag(kMaskGpuTag)) {
+    cc->Inputs().Tag(kMaskGpuTag).Set<mediapipe::GpuBuffer>();
     use_gpu |= true;
   }
 #endif  //  !MEDIAPIPE_DISABLE_GPU
-  if (cc->Inputs().HasTag("MASK")) {
-    cc->Inputs().Tag("MASK").Set<ImageFrame>();
+  if (cc->Inputs().HasTag(kMaskCpuTag)) {
+    cc->Inputs().Tag(kMaskCpuTag).Set<ImageFrame>();
   }
 
 #if !defined(MEDIAPIPE_DISABLE_GPU)
-  if (cc->Outputs().HasTag("IMAGE_GPU")) {
-    cc->Outputs().Tag("IMAGE_GPU").Set<mediapipe::GpuBuffer>();
+  if (cc->Outputs().HasTag(kGpuBufferTag)) {
+    cc->Outputs().Tag(kGpuBufferTag).Set<mediapipe::GpuBuffer>();
     use_gpu |= true;
   }
 #endif  //  !MEDIAPIPE_DISABLE_GPU
-  if (cc->Outputs().HasTag("IMAGE")) {
-    cc->Outputs().Tag("IMAGE").Set<ImageFrame>();
+  if (cc->Outputs().HasTag(kImageFrameTag)) {
+    cc->Outputs().Tag(kImageFrameTag).Set<ImageFrame>();
   }
 
   // Confirm only one of the input streams is present.
-  RET_CHECK(cc->Inputs().HasTag("IMAGE") ^ cc->Inputs().HasTag("IMAGE_GPU"));
+  RET_CHECK(cc->Inputs().HasTag(kImageFrameTag) ^
+            cc->Inputs().HasTag(kGpuBufferTag));
   // Confirm only one of the output streams is present.
-  RET_CHECK(cc->Outputs().HasTag("IMAGE") ^ cc->Outputs().HasTag("IMAGE_GPU"));
+  RET_CHECK(cc->Outputs().HasTag(kImageFrameTag) ^
+            cc->Outputs().HasTag(kGpuBufferTag));
 
   if (use_gpu) {
 #if !defined(MEDIAPIPE_DISABLE_GPU)
@@ -158,7 +165,7 @@ REGISTER_CALCULATOR(RecolorCalculator);
 ::mediapipe::Status RecolorCalculator::Open(CalculatorContext* cc) {
   cc->SetOffset(TimestampDiff(0));
 
-  if (cc->Inputs().HasTag("IMAGE_GPU")) {
+  if (cc->Inputs().HasTag(kGpuBufferTag)) {
     use_gpu_ = true;
 #if !defined(MEDIAPIPE_DISABLE_GPU)
     MP_RETURN_IF_ERROR(gpu_helper_.Open(cc));
@@ -201,12 +208,12 @@ REGISTER_CALCULATOR(RecolorCalculator);
 }
 
 ::mediapipe::Status RecolorCalculator::RenderCpu(CalculatorContext* cc) {
-  if (cc->Inputs().Tag("MASK").IsEmpty()) {
+  if (cc->Inputs().Tag(kMaskCpuTag).IsEmpty()) {
     return ::mediapipe::OkStatus();
   }
   // Get inputs and setup output.
-  const auto& input_img = cc->Inputs().Tag("IMAGE").Get<ImageFrame>();
-  const auto& mask_img = cc->Inputs().Tag("MASK").Get<ImageFrame>();
+  const auto& input_img = cc->Inputs().Tag(kImageFrameTag).Get<ImageFrame>();
+  const auto& mask_img = cc->Inputs().Tag(kMaskCpuTag).Get<ImageFrame>();
 
   cv::Mat input_mat = formats::MatView(&input_img);
   cv::Mat mask_mat = formats::MatView(&mask_img);
@@ -254,19 +261,21 @@ REGISTER_CALCULATOR(RecolorCalculator);
     }
   }
 
-  cc->Outputs().Tag("IMAGE").Add(output_img.release(), cc->InputTimestamp());
+  cc->Outputs()
+      .Tag(kImageFrameTag)
+      .Add(output_img.release(), cc->InputTimestamp());
 
   return ::mediapipe::OkStatus();
 }
 
 ::mediapipe::Status RecolorCalculator::RenderGpu(CalculatorContext* cc) {
-  if (cc->Inputs().Tag("MASK_GPU").IsEmpty()) {
+  if (cc->Inputs().Tag(kMaskGpuTag).IsEmpty()) {
     return ::mediapipe::OkStatus();
   }
 #if !defined(MEDIAPIPE_DISABLE_GPU)
   // Get inputs and setup output.
-  const Packet& input_packet = cc->Inputs().Tag("IMAGE_GPU").Value();
-  const Packet& mask_packet = cc->Inputs().Tag("MASK_GPU").Value();
+  const Packet& input_packet = cc->Inputs().Tag(kGpuBufferTag).Value();
+  const Packet& mask_packet = cc->Inputs().Tag(kMaskGpuTag).Value();
 
   const auto& input_buffer = input_packet.Get<mediapipe::GpuBuffer>();
   const auto& mask_buffer = mask_packet.Get<mediapipe::GpuBuffer>();
@@ -296,7 +305,7 @@ REGISTER_CALCULATOR(RecolorCalculator);
 
   // Send result image in GPU packet.
   auto output = dst_tex.GetFrame<mediapipe::GpuBuffer>();
-  cc->Outputs().Tag("IMAGE_GPU").Add(output.release(), cc->InputTimestamp());
+  cc->Outputs().Tag(kGpuBufferTag).Add(output.release(), cc->InputTimestamp());
 
   // Cleanup
   img_tex.Release();

mediapipe/calculators/tflite/BUILD

@@ -243,6 +243,7 @@ cc_library(
             "@org_tensorflow//tensorflow/lite/delegates/gpu:metal_delegate_internal",
         ],
         "//conditions:default": [
+            "//mediapipe/util/tflite:tflite_gpu_runner",
             "//mediapipe/gpu:gl_calculator_helper",
             "//mediapipe/gpu:gpu_buffer",
             "@org_tensorflow//tensorflow/lite/delegates/gpu/common:shape",

mediapipe/calculators/tflite/tflite_converter_calculator.cc

@@ -63,6 +63,10 @@ typedef Eigen::Matrix<float, Eigen::Dynamic, Eigen::Dynamic, Eigen::RowMajor>
 typedef Eigen::Matrix<float, Eigen::Dynamic, Eigen::Dynamic, Eigen::ColMajor>
     ColMajorMatrixXf;
 
+constexpr char kImageFrameTag[] = "IMAGE";
+constexpr char kGpuBufferTag[] = "IMAGE_GPU";
+constexpr char kTensorsTag[] = "TENSORS";
+constexpr char kTensorsGpuTag[] = "TENSORS_GPU";
 }  // namespace
 
 namespace mediapipe {
@@ -124,6 +128,9 @@ struct GPUData {
 //  GPU tensors are currently only supported on mobile platforms.
 //  This calculator uses FixedSizeInputStreamHandler by default.
 //
+// Note: Input defines output, so only these type sets are supported:
+// IMAGE -> TENSORS | IMAGE_GPU -> TENSORS_GPU | MATRIX -> TENSORS
+//
 class TfLiteConverterCalculator : public CalculatorBase {
  public:
   static ::mediapipe::Status GetContract(CalculatorContract* cc);
@@ -138,9 +145,9 @@ class TfLiteConverterCalculator : public CalculatorBase {
   template <class T>
   ::mediapipe::Status NormalizeImage(const ImageFrame& image_frame,
                                      bool zero_center, bool flip_vertically,
-                                     float* tensor_buffer);
+                                     float* tensor_ptr);
   ::mediapipe::Status CopyMatrixToTensor(const Matrix& matrix,
-                                         float* tensor_buffer);
+                                         float* tensor_ptr);
   ::mediapipe::Status ProcessCPU(CalculatorContext* cc);
   ::mediapipe::Status ProcessGPU(CalculatorContext* cc);
 
@@ -166,33 +173,35 @@ REGISTER_CALCULATOR(TfLiteConverterCalculator);
 
 ::mediapipe::Status TfLiteConverterCalculator::GetContract(
     CalculatorContract* cc) {
-  const bool has_image_tag = cc->Inputs().HasTag("IMAGE");
-  const bool has_image_gpu_tag = cc->Inputs().HasTag("IMAGE_GPU");
-  const bool has_matrix_tag = cc->Inputs().HasTag("MATRIX");
   // Confirm only one of the input streams is present.
-  RET_CHECK(has_image_tag ^ has_image_gpu_tag ^ has_matrix_tag &&
-            !(has_image_tag && has_image_gpu_tag && has_matrix_tag));
+  RET_CHECK(cc->Inputs().HasTag(kImageFrameTag) ^
+            cc->Inputs().HasTag(kGpuBufferTag) ^ cc->Inputs().HasTag("MATRIX"));
 
   // Confirm only one of the output streams is present.
-  RET_CHECK(cc->Outputs().HasTag("TENSORS") ^
-            cc->Outputs().HasTag("TENSORS_GPU"));
+  RET_CHECK(cc->Outputs().HasTag(kTensorsTag) ^
+            cc->Outputs().HasTag(kTensorsGpuTag));
 
   bool use_gpu = false;
 
-  if (cc->Inputs().HasTag("IMAGE")) cc->Inputs().Tag("IMAGE").Set<ImageFrame>();
-  if (cc->Inputs().HasTag("MATRIX")) cc->Inputs().Tag("MATRIX").Set<Matrix>();
+  if (cc->Inputs().HasTag(kImageFrameTag)) {
+    cc->Inputs().Tag(kImageFrameTag).Set<ImageFrame>();
+  }
+  if (cc->Inputs().HasTag("MATRIX")) {
+    cc->Inputs().Tag("MATRIX").Set<Matrix>();
+  }
 #if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__)
-  if (cc->Inputs().HasTag("IMAGE_GPU")) {
-    cc->Inputs().Tag("IMAGE_GPU").Set<mediapipe::GpuBuffer>();
+  if (cc->Inputs().HasTag(kGpuBufferTag)) {
+    cc->Inputs().Tag(kGpuBufferTag).Set<mediapipe::GpuBuffer>();
     use_gpu |= true;
   }
 #endif  //  !MEDIAPIPE_DISABLE_GPU
 
-  if (cc->Outputs().HasTag("TENSORS"))
-    cc->Outputs().Tag("TENSORS").Set<std::vector<TfLiteTensor>>();
+  if (cc->Outputs().HasTag(kTensorsTag)) {
+    cc->Outputs().Tag(kTensorsTag).Set<std::vector<TfLiteTensor>>();
+  }
 #if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__)
-  if (cc->Outputs().HasTag("TENSORS_GPU")) {
-    cc->Outputs().Tag("TENSORS_GPU").Set<std::vector<GpuTensor>>();
+  if (cc->Outputs().HasTag(kTensorsGpuTag)) {
+    cc->Outputs().Tag(kTensorsGpuTag).Set<std::vector<GpuTensor>>();
     use_gpu |= true;
   }
 #endif  //  !MEDIAPIPE_DISABLE_GPU
@@ -216,8 +225,8 @@ REGISTER_CALCULATOR(TfLiteConverterCalculator);
 
   MP_RETURN_IF_ERROR(LoadOptions(cc));
 
-  if (cc->Inputs().HasTag("IMAGE_GPU") ||
-      cc->Outputs().HasTag("IMAGE_OUT_GPU")) {
+  if (cc->Inputs().HasTag(kGpuBufferTag) ||
+      cc->Outputs().HasTag(kGpuBufferTag)) {
 #if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__)
     use_gpu_ = true;
 #else
@@ -227,8 +236,8 @@ REGISTER_CALCULATOR(TfLiteConverterCalculator);
 
   if (use_gpu_) {
     // Cannot mix CPU/GPU streams.
-    RET_CHECK(cc->Inputs().HasTag("IMAGE_GPU") &&
-              cc->Outputs().HasTag("TENSORS_GPU"));
+    RET_CHECK(cc->Inputs().HasTag(kGpuBufferTag) &&
+              cc->Outputs().HasTag(kTensorsGpuTag));
     // Cannot use quantization.
     use_quantized_tensors_ = false;
 #if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
@@ -248,7 +257,6 @@ REGISTER_CALCULATOR(TfLiteConverterCalculator);
 
 ::mediapipe::Status TfLiteConverterCalculator::Process(CalculatorContext* cc) {
   if (use_gpu_) {
-    // GpuBuffer to tflite::gpu::GlBuffer conversion.
     if (!initialized_) {
       MP_RETURN_IF_ERROR(InitGpu(cc));
       initialized_ = true;
@@ -259,7 +267,6 @@ REGISTER_CALCULATOR(TfLiteConverterCalculator);
     // Convert to CPU tensors or Matrix type.
     MP_RETURN_IF_ERROR(ProcessCPU(cc));
   }
-
   return ::mediapipe::OkStatus();
 }
 
@@ -275,24 +282,26 @@ REGISTER_CALCULATOR(TfLiteConverterCalculator);
 
 ::mediapipe::Status TfLiteConverterCalculator::ProcessCPU(
     CalculatorContext* cc) {
-  if (cc->Inputs().HasTag("IMAGE")) {
+  if (cc->Inputs().HasTag(kImageFrameTag)) {
     // CPU ImageFrame to TfLiteTensor conversion.
 
-    const auto& image_frame = cc->Inputs().Tag("IMAGE").Get<ImageFrame>();
+    const auto& image_frame =
+        cc->Inputs().Tag(kImageFrameTag).Get<ImageFrame>();
     const int height = image_frame.Height();
     const int width = image_frame.Width();
     const int channels = image_frame.NumberOfChannels();
     const int channels_preserved = std::min(channels, max_num_channels_);
+    const mediapipe::ImageFormat::Format format = image_frame.Format();
 
     if (!initialized_) {
-      if (!(image_frame.Format() == mediapipe::ImageFormat::SRGBA ||
-            image_frame.Format() == mediapipe::ImageFormat::SRGB ||
-            image_frame.Format() == mediapipe::ImageFormat::GRAY8 ||
-            image_frame.Format() == mediapipe::ImageFormat::VEC32F1))
+      if (!(format == mediapipe::ImageFormat::SRGBA ||
+            format == mediapipe::ImageFormat::SRGB ||
+            format == mediapipe::ImageFormat::GRAY8 ||
+            format == mediapipe::ImageFormat::VEC32F1))
         RET_CHECK_FAIL() << "Unsupported CPU input format.";
       TfLiteQuantization quant;
       if (use_quantized_tensors_) {
-        RET_CHECK(image_frame.Format() != mediapipe::ImageFormat::VEC32F1)
+        RET_CHECK(format != mediapipe::ImageFormat::VEC32F1)
             << "Only 8-bit input images are supported for quantization.";
         quant.type = kTfLiteAffineQuantization;
         quant.params = nullptr;
@@ -349,8 +358,9 @@ REGISTER_CALCULATOR(TfLiteConverterCalculator);
 
     auto output_tensors = absl::make_unique<std::vector<TfLiteTensor>>();
     output_tensors->emplace_back(*tensor);
-    cc->Outputs().Tag("TENSORS").Add(output_tensors.release(),
-                                     cc->InputTimestamp());
+    cc->Outputs()
+        .Tag(kTensorsTag)
+        .Add(output_tensors.release(), cc->InputTimestamp());
   } else if (cc->Inputs().HasTag("MATRIX")) {
     // CPU Matrix to TfLiteTensor conversion.
 
@@ -371,15 +381,16 @@ REGISTER_CALCULATOR(TfLiteConverterCalculator);
     interpreter_->ResizeInputTensor(tensor_idx, {height, width, channels});
     interpreter_->AllocateTensors();
 
-    float* tensor_buffer = tensor->data.f;
-    RET_CHECK(tensor_buffer);
+    float* tensor_ptr = tensor->data.f;
+    RET_CHECK(tensor_ptr);
 
-    MP_RETURN_IF_ERROR(CopyMatrixToTensor(matrix, tensor_buffer));
+    MP_RETURN_IF_ERROR(CopyMatrixToTensor(matrix, tensor_ptr));
 
     auto output_tensors = absl::make_unique<std::vector<TfLiteTensor>>();
     output_tensors->emplace_back(*tensor);
-    cc->Outputs().Tag("TENSORS").Add(output_tensors.release(),
-                                     cc->InputTimestamp());
+    cc->Outputs()
+        .Tag(kTensorsTag)
+        .Add(output_tensors.release(), cc->InputTimestamp());
   }
 
   return ::mediapipe::OkStatus();
@@ -389,7 +400,8 @@ REGISTER_CALCULATOR(TfLiteConverterCalculator);
     CalculatorContext* cc) {
 #if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
   // GpuBuffer to tflite::gpu::GlBuffer conversion.
-  const auto& input = cc->Inputs().Tag("IMAGE_GPU").Get<mediapipe::GpuBuffer>();
+  const auto& input =
+      cc->Inputs().Tag(kGpuBufferTag).Get<mediapipe::GpuBuffer>();
   MP_RETURN_IF_ERROR(
       gpu_helper_.RunInGlContext([this, &input]() -> ::mediapipe::Status {
         // Convert GL texture into TfLite GlBuffer (SSBO).
@@ -421,11 +433,12 @@ REGISTER_CALCULATOR(TfLiteConverterCalculator);
         return ::mediapipe::OkStatus();
       }));
   cc->Outputs()
-      .Tag("TENSORS_GPU")
+      .Tag(kTensorsGpuTag)
       .Add(output_tensors.release(), cc->InputTimestamp());
 #elif defined(MEDIAPIPE_IOS)
   // GpuBuffer to id<MTLBuffer> conversion.
-  const auto& input = cc->Inputs().Tag("IMAGE_GPU").Get<mediapipe::GpuBuffer>();
+  const auto& input =
+      cc->Inputs().Tag(kGpuBufferTag).Get<mediapipe::GpuBuffer>();
   id<MTLCommandBuffer> command_buffer = [gpu_helper_ commandBuffer];
 
   id<MTLTexture> src_texture = [gpu_helper_ metalTextureWithGpuBuffer:input];
@@ -457,7 +470,7 @@ REGISTER_CALCULATOR(TfLiteConverterCalculator);
                     commandBuffer:command_buffer];
 
   cc->Outputs()
-      .Tag("TENSORS_GPU")
+      .Tag(kTensorsGpuTag)
       .Add(output_tensors.release(), cc->InputTimestamp());
 #else
   RET_CHECK_FAIL() << "GPU processing is not enabled.";
@@ -469,7 +482,8 @@ REGISTER_CALCULATOR(TfLiteConverterCalculator);
 ::mediapipe::Status TfLiteConverterCalculator::InitGpu(CalculatorContext* cc) {
 #if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__)
   // Get input image sizes.
-  const auto& input = cc->Inputs().Tag("IMAGE_GPU").Get<mediapipe::GpuBuffer>();
+  const auto& input =
+      cc->Inputs().Tag(kGpuBufferTag).Get<mediapipe::GpuBuffer>();
   mediapipe::ImageFormat::Format format =
       mediapipe::ImageFormatForGpuBufferFormat(input.format());
   gpu_data_out_ = absl::make_unique<GPUData>();
@@ -612,7 +626,7 @@ REGISTER_CALCULATOR(TfLiteConverterCalculator);
   CHECK_LE(max_num_channels_, 4);
   CHECK_NE(max_num_channels_, 2);
 #if defined(MEDIAPIPE_IOS)
-  if (cc->Inputs().HasTag("IMAGE_GPU"))
+  if (cc->Inputs().HasTag(kGpuBufferTag))
     // Currently on iOS, tflite gpu input tensor must be 4 channels,
     // so input image must be 4 channels also (checked in InitGpu).
     max_num_channels_ = 4;
@@ -627,7 +641,7 @@ REGISTER_CALCULATOR(TfLiteConverterCalculator);
 template <class T>
 ::mediapipe::Status TfLiteConverterCalculator::NormalizeImage(
     const ImageFrame& image_frame, bool zero_center, bool flip_vertically,
-    float* tensor_buffer) {
+    float* tensor_ptr) {
   const int height = image_frame.Height();
   const int width = image_frame.Width();
   const int channels = image_frame.NumberOfChannels();
@@ -651,7 +665,7 @@ template <class T>
         (flip_vertically ? height - 1 - i : i) * image_frame.WidthStep());
     for (int j = 0; j < width; ++j) {
       for (int c = 0; c < channels_preserved; ++c) {
-        *tensor_buffer++ = *image_ptr++ / div - sub;
+        *tensor_ptr++ = *image_ptr++ / div - sub;
       }
       image_ptr += channels_ignored;
     }
@@ -661,14 +675,14 @@ template <class T>
 }
 
 ::mediapipe::Status TfLiteConverterCalculator::CopyMatrixToTensor(
-    const Matrix& matrix, float* tensor_buffer) {
+    const Matrix& matrix, float* tensor_ptr) {
   if (row_major_matrix_) {
-    auto matrix_map = Eigen::Map<RowMajorMatrixXf>(tensor_buffer, matrix.rows(),
-                                                   matrix.cols());
+    auto matrix_map =
+        Eigen::Map<RowMajorMatrixXf>(tensor_ptr, matrix.rows(), matrix.cols());
     matrix_map = matrix;
   } else {
-    auto matrix_map = Eigen::Map<ColMajorMatrixXf>(tensor_buffer, matrix.rows(),
-                                                   matrix.cols());
+    auto matrix_map =
+        Eigen::Map<ColMajorMatrixXf>(tensor_ptr, matrix.rows(), matrix.cols());
     matrix_map = matrix;
   }
 

mediapipe/calculators/tflite/tflite_inference_calculator.cc

@@ -36,6 +36,7 @@
 #if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
 #include "mediapipe/gpu/gl_calculator_helper.h"
 #include "mediapipe/gpu/gpu_buffer.h"
+#include "mediapipe/util/tflite/tflite_gpu_runner.h"
 #include "tensorflow/lite/delegates/gpu/common/shape.h"
 #include "tensorflow/lite/delegates/gpu/gl/gl_buffer.h"
 #include "tensorflow/lite/delegates/gpu/gl/gl_program.h"
@@ -75,6 +76,9 @@ typedef id<MTLBuffer> GpuTensor;
 
 // Round up n to next multiple of m.
 size_t RoundUp(size_t n, size_t m) { return ((n + m - 1) / m) * m; }  // NOLINT
+
+constexpr char kTensorsTag[] = "TENSORS";
+constexpr char kTensorsGpuTag[] = "TENSORS_GPU";
 }  // namespace
 
 #if defined(MEDIAPIPE_EDGE_TPU)
@@ -219,6 +223,7 @@ class TfLiteInferenceCalculator : public CalculatorBase {
   ::mediapipe::Status LoadModel(CalculatorContext* cc);
   ::mediapipe::StatusOr<Packet> GetModelAsPacket(const CalculatorContext& cc);
   ::mediapipe::Status LoadDelegate(CalculatorContext* cc);
+  ::mediapipe::Status InitTFLiteGPURunner();
 
   Packet model_packet_;
   std::unique_ptr<tflite::Interpreter> interpreter_;
@@ -228,6 +233,7 @@ class TfLiteInferenceCalculator : public CalculatorBase {
   mediapipe::GlCalculatorHelper gpu_helper_;
   std::vector<std::unique_ptr<GPUData>> gpu_data_in_;
   std::vector<std::unique_ptr<GPUData>> gpu_data_out_;
+  std::unique_ptr<tflite::gpu::TFLiteGPURunner> tflite_gpu_runner_;
 #elif defined(MEDIAPIPE_IOS)
   MPPMetalHelper* gpu_helper_ = nullptr;
   std::vector<std::unique_ptr<GPUData>> gpu_data_in_;
@@ -245,6 +251,8 @@ class TfLiteInferenceCalculator : public CalculatorBase {
   bool gpu_input_ = false;
   bool gpu_output_ = false;
   bool use_quantized_tensors_ = false;
+
+  bool use_advanced_gpu_api_ = false;
 };
 REGISTER_CALCULATOR(TfLiteInferenceCalculator);
 
@@ -252,10 +260,10 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
 
 ::mediapipe::Status TfLiteInferenceCalculator::GetContract(
     CalculatorContract* cc) {
-  RET_CHECK(cc->Inputs().HasTag("TENSORS") ^
-            cc->Inputs().HasTag("TENSORS_GPU"));
-  RET_CHECK(cc->Outputs().HasTag("TENSORS") ^
-            cc->Outputs().HasTag("TENSORS_GPU"));
+  RET_CHECK(cc->Inputs().HasTag(kTensorsTag) ^
+            cc->Inputs().HasTag(kTensorsGpuTag));
+  RET_CHECK(cc->Outputs().HasTag(kTensorsTag) ^
+            cc->Outputs().HasTag(kTensorsGpuTag));
 
   const auto& options =
       cc->Options<::mediapipe::TfLiteInferenceCalculatorOptions>();
@@ -266,26 +274,26 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
   bool use_gpu =
       options.has_delegate() ? options.delegate().has_gpu() : options.use_gpu();
 
-  if (cc->Inputs().HasTag("TENSORS"))
-    cc->Inputs().Tag("TENSORS").Set<std::vector<TfLiteTensor>>();
+  if (cc->Inputs().HasTag(kTensorsTag))
+    cc->Inputs().Tag(kTensorsTag).Set<std::vector<TfLiteTensor>>();
 #if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__)
-  if (cc->Inputs().HasTag("TENSORS_GPU")) {
+  if (cc->Inputs().HasTag(kTensorsGpuTag)) {
     RET_CHECK(!options.has_delegate() || options.delegate().has_gpu())
         << "GPU input is compatible with GPU delegate only.";
 
-    cc->Inputs().Tag("TENSORS_GPU").Set<std::vector<GpuTensor>>();
+    cc->Inputs().Tag(kTensorsGpuTag).Set<std::vector<GpuTensor>>();
     use_gpu |= true;
   }
 #endif  //  !MEDIAPIPE_DISABLE_GPU
 
-  if (cc->Outputs().HasTag("TENSORS"))
-    cc->Outputs().Tag("TENSORS").Set<std::vector<TfLiteTensor>>();
+  if (cc->Outputs().HasTag(kTensorsTag))
+    cc->Outputs().Tag(kTensorsTag).Set<std::vector<TfLiteTensor>>();
 #if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__)
-  if (cc->Outputs().HasTag("TENSORS_GPU")) {
+  if (cc->Outputs().HasTag(kTensorsGpuTag)) {
     RET_CHECK(!options.has_delegate() || options.delegate().has_gpu())
         << "GPU output is compatible with GPU delegate only.";
 
-    cc->Outputs().Tag("TENSORS_GPU").Set<std::vector<GpuTensor>>();
+    cc->Outputs().Tag(kTensorsGpuTag).Set<std::vector<GpuTensor>>();
     use_gpu |= true;
   }
 #endif  //  !MEDIAPIPE_DISABLE_GPU
@@ -320,27 +328,31 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
       cc->Options<::mediapipe::TfLiteInferenceCalculatorOptions>();
   gpu_inference_ = options.use_gpu();
 
-  if (cc->Inputs().HasTag("TENSORS_GPU")) {
+  if (cc->Inputs().HasTag(kTensorsGpuTag)) {
 #if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__)
     gpu_input_ = true;
     gpu_inference_ = true;  // Inference must be on GPU also.
 #else
-    RET_CHECK(!cc->Inputs().HasTag("TENSORS_GPU"))
+    RET_CHECK(!cc->Inputs().HasTag(kTensorsGpuTag))
         << "GPU processing not enabled.";
 #endif  //  !MEDIAPIPE_DISABLE_GPU
   }
 
-  if (cc->Outputs().HasTag("TENSORS_GPU")) {
+  if (cc->Outputs().HasTag(kTensorsGpuTag)) {
 #if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__)
     gpu_output_ = true;
-    RET_CHECK(cc->Inputs().HasTag("TENSORS_GPU"))
+    RET_CHECK(cc->Inputs().HasTag(kTensorsGpuTag))
         << "GPU output must also have GPU Input.";
 #else
-    RET_CHECK(!cc->Inputs().HasTag("TENSORS_GPU"))
+    RET_CHECK(!cc->Inputs().HasTag(kTensorsGpuTag))
         << "GPU processing not enabled.";
 #endif  //  !MEDIAPIPE_DISABLE_GPU
   }
 
+  const auto& calculator_opts =
+      cc->Options<mediapipe::TfLiteInferenceCalculatorOptions>();
+  use_advanced_gpu_api_ = false;
+
   MP_RETURN_IF_ERROR(LoadModel(cc));
 
   if (gpu_inference_) {
@@ -352,8 +364,12 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
 #endif
 
 #if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
-    MP_RETURN_IF_ERROR(gpu_helper_.RunInGlContext(
-        [this, &cc]() -> ::mediapipe::Status { return LoadDelegate(cc); }));
+    MP_RETURN_IF_ERROR(
+        gpu_helper_.RunInGlContext([this, &cc]() -> ::mediapipe::Status {
+          return use_advanced_gpu_api_ ? InitTFLiteGPURunner()
+                                       : LoadDelegate(cc);
+        }));
+    if (use_advanced_gpu_api_) return ::mediapipe::OkStatus();
 #else
     MP_RETURN_IF_ERROR(LoadDelegate(cc));
 #endif
@@ -365,13 +381,51 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
   return ::mediapipe::OkStatus();
 }
 
+::mediapipe::Status TfLiteInferenceCalculator::InitTFLiteGPURunner() {
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
+  // Create and bind OpenGL buffers for outputs.
+  // These buffers are created onve and later their ids are jut passed to the
+  // calculator outputs.
+
+  gpu_data_out_.resize(tflite_gpu_runner_->outputs_size());
+  for (int i = 0; i < tflite_gpu_runner_->outputs_size(); ++i) {
+    gpu_data_out_[i] = absl::make_unique<GPUData>();
+    ASSIGN_OR_RETURN(gpu_data_out_[i]->elements,
+                     tflite_gpu_runner_->GetOutputElements(i));
+    // Create and bind input buffer.
+    RET_CHECK_CALL(::tflite::gpu::gl::CreateReadWriteShaderStorageBuffer<float>(
+        gpu_data_out_[i]->elements, &gpu_data_out_[i]->buffer));
+  }
+  RET_CHECK_CALL(tflite_gpu_runner_->Build());
+#endif
+  return ::mediapipe::OkStatus();
+}
+
 ::mediapipe::Status TfLiteInferenceCalculator::Process(CalculatorContext* cc) {
   // 1. Receive pre-processed tensor inputs.
-  if (gpu_input_) {
-    // Read GPU input into SSBO.
+  if (use_advanced_gpu_api_) {
 #if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
     const auto& input_tensors =
         cc->Inputs().Tag("TENSORS_GPU").Get<std::vector<GpuTensor>>();
+    RET_CHECK(input_tensors.empty());
+    MP_RETURN_IF_ERROR(gpu_helper_.RunInGlContext(
+        [this, &input_tensors]() -> ::mediapipe::Status {
+          for (int i = 0; i < input_tensors.size(); ++i) {
+            MP_RETURN_IF_ERROR(tflite_gpu_runner_->BindSSBOToInputTensor(
+                input_tensors[i].id(), i));
+          }
+          for (int i = 0; i < gpu_data_out_.size(); ++i) {
+            MP_RETURN_IF_ERROR(tflite_gpu_runner_->BindSSBOToOutputTensor(
+                gpu_data_out_[i]->buffer.id(), i));
+          }
+          return ::mediapipe::OkStatus();
+        }));
+#endif
+  } else if (gpu_input_) {
+    // Read GPU input into SSBO.
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
+    const auto& input_tensors =
+        cc->Inputs().Tag(kTensorsGpuTag).Get<std::vector<GpuTensor>>();
     RET_CHECK_GT(input_tensors.size(), 0);
     MP_RETURN_IF_ERROR(gpu_helper_.RunInGlContext(
         [this, &input_tensors]() -> ::mediapipe::Status {
@@ -386,7 +440,7 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
         }));
 #elif defined(MEDIAPIPE_IOS)
     const auto& input_tensors =
-        cc->Inputs().Tag("TENSORS_GPU").Get<std::vector<GpuTensor>>();
+        cc->Inputs().Tag(kTensorsGpuTag).Get<std::vector<GpuTensor>>();
     RET_CHECK_GT(input_tensors.size(), 0);
     // Explicit copy input with conversion float 32 bits to 16 bits.
     gpu_data_in_.resize(input_tensors.size());
@@ -413,7 +467,7 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
   } else {
     // Read CPU input into tensors.
     const auto& input_tensors =
-        cc->Inputs().Tag("TENSORS").Get<std::vector<TfLiteTensor>>();
+        cc->Inputs().Tag(kTensorsTag).Get<std::vector<TfLiteTensor>>();
     RET_CHECK_GT(input_tensors.size(), 0);
     for (int i = 0; i < input_tensors.size(); ++i) {
       const TfLiteTensor* input_tensor = &input_tensors[i];
@@ -437,7 +491,11 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
 #if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
     MP_RETURN_IF_ERROR(
         gpu_helper_.RunInGlContext([this]() -> ::mediapipe::Status {
-          RET_CHECK_EQ(interpreter_->Invoke(), kTfLiteOk);
+          if (use_advanced_gpu_api_) {
+            RET_CHECK(tflite_gpu_runner_->Invoke().ok());
+          } else {
+            RET_CHECK_EQ(interpreter_->Invoke(), kTfLiteOk);
+          }
           return ::mediapipe::OkStatus();
         }));
 #elif defined(MEDIAPIPE_IOS)
@@ -448,7 +506,18 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
   }
 
   // 3. Output processed tensors.
-  if (gpu_output_) {
+  if (use_advanced_gpu_api_) {
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
+    auto output_tensors = absl::make_unique<std::vector<GpuTensor>>();
+    output_tensors->resize(gpu_data_out_.size());
+    for (int i = 0; i < gpu_data_out_.size(); ++i) {
+      output_tensors->at(i) = gpu_data_out_[0]->buffer.MakeRef();
+    }
+    cc->Outputs()
+        .Tag("TENSORS_GPU")
+        .Add(output_tensors.release(), cc->InputTimestamp());
+#endif
+  } else if (gpu_output_) {
 #if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
     // Output result tensors (GPU).
     auto output_tensors = absl::make_unique<std::vector<GpuTensor>>();
@@ -464,7 +533,7 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
           return ::mediapipe::OkStatus();
         }));
     cc->Outputs()
-        .Tag("TENSORS_GPU")
+        .Tag(kTensorsGpuTag)
         .Add(output_tensors.release(), cc->InputTimestamp());
 #elif defined(MEDIAPIPE_IOS)
     // Output result tensors (GPU).
@@ -488,7 +557,7 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
     [convert_command endEncoding];
     [command_buffer commit];
     cc->Outputs()
-        .Tag("TENSORS_GPU")
+        .Tag(kTensorsGpuTag)
         .Add(output_tensors.release(), cc->InputTimestamp());
 #else
     RET_CHECK_FAIL() << "GPU processing not enabled.";
@@ -501,8 +570,9 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
       TfLiteTensor* tensor = interpreter_->tensor(tensor_indexes[i]);
       output_tensors->emplace_back(*tensor);
     }
-    cc->Outputs().Tag("TENSORS").Add(output_tensors.release(),
-                                     cc->InputTimestamp());
+    cc->Outputs()
+        .Tag(kTensorsTag)
+        .Add(output_tensors.release(), cc->InputTimestamp());
   }
 
   return ::mediapipe::OkStatus();
@@ -557,6 +627,20 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
                       .Tag("CUSTOM_OP_RESOLVER")
                       .Get<tflite::ops::builtin::BuiltinOpResolver>();
   }
+
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
+  if (use_advanced_gpu_api_) {
+    tflite::gpu::InferenceOptions options;
+    options.priority1 = tflite::gpu::InferencePriority::MIN_LATENCY;
+    options.priority2 = tflite::gpu::InferencePriority::AUTO;
+    options.priority3 = tflite::gpu::InferencePriority::AUTO;
+    options.usage = tflite::gpu::InferenceUsage::SUSTAINED_SPEED;
+    tflite_gpu_runner_ =
+        std::make_unique<tflite::gpu::TFLiteGPURunner>(options);
+    return tflite_gpu_runner_->InitializeWithModel(model);
+  }
+#endif
+
 #if defined(MEDIAPIPE_EDGE_TPU)
   interpreter_ =
       BuildEdgeTpuInterpreter(model, &op_resolver, edgetpu_context_.get());

mediapipe/calculators/tflite/tflite_inference_calculator.proto

@@ -42,7 +42,11 @@ message TfLiteInferenceCalculatorOptions {
     message TfLite {}
     // Delegate to run GPU inference depending on the device.
     // (Can use OpenGl, OpenCl, Metal depending on the device.)
-    message Gpu {}
+    message Gpu {
+      // Experimental, Android/Linux only. Use TFLite GPU delegate API2 for
+      // the NN inference.
+      optional bool use_advanced_gpu_api = 1 [default = false];
+    }
     // Android only.
     message Nnapi {}
     message Xnnpack {

mediapipe/calculators/tflite/tflite_tensors_to_detections_calculator.cc

@@ -47,10 +47,11 @@
 #endif  // iOS
 
 namespace {
-
 constexpr int kNumInputTensorsWithAnchors = 3;
 constexpr int kNumCoordsPerBox = 4;
 
+constexpr char kTensorsTag[] = "TENSORS";
+constexpr char kTensorsGpuTag[] = "TENSORS_GPU";
 }  // namespace
 
 namespace mediapipe {
@@ -200,13 +201,13 @@ REGISTER_CALCULATOR(TfLiteTensorsToDetectionsCalculator);
 
   bool use_gpu = false;
 
-  if (cc->Inputs().HasTag("TENSORS")) {
-    cc->Inputs().Tag("TENSORS").Set<std::vector<TfLiteTensor>>();
+  if (cc->Inputs().HasTag(kTensorsTag)) {
+    cc->Inputs().Tag(kTensorsTag).Set<std::vector<TfLiteTensor>>();
   }
 
 #if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__)
-  if (cc->Inputs().HasTag("TENSORS_GPU")) {
-    cc->Inputs().Tag("TENSORS_GPU").Set<std::vector<GpuTensor>>();
+  if (cc->Inputs().HasTag(kTensorsGpuTag)) {
+    cc->Inputs().Tag(kTensorsGpuTag).Set<std::vector<GpuTensor>>();
     use_gpu |= true;
   }
 #endif  //  !MEDIAPIPE_DISABLE_GPU
@@ -236,7 +237,7 @@ REGISTER_CALCULATOR(TfLiteTensorsToDetectionsCalculator);
     CalculatorContext* cc) {
   cc->SetOffset(TimestampDiff(0));
 
-  if (cc->Inputs().HasTag("TENSORS_GPU")) {
+  if (cc->Inputs().HasTag(kTensorsGpuTag)) {
     gpu_input_ = true;
 #if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
     MP_RETURN_IF_ERROR(gpu_helper_.Open(cc));
@@ -258,8 +259,8 @@ REGISTER_CALCULATOR(TfLiteTensorsToDetectionsCalculator);
 
 ::mediapipe::Status TfLiteTensorsToDetectionsCalculator::Process(
     CalculatorContext* cc) {
-  if ((!gpu_input_ && cc->Inputs().Tag("TENSORS").IsEmpty()) ||
-      (gpu_input_ && cc->Inputs().Tag("TENSORS_GPU").IsEmpty())) {
+  if ((!gpu_input_ && cc->Inputs().Tag(kTensorsTag).IsEmpty()) ||
+      (gpu_input_ && cc->Inputs().Tag(kTensorsGpuTag).IsEmpty())) {
     return ::mediapipe::OkStatus();
   }
 
@@ -284,7 +285,7 @@ REGISTER_CALCULATOR(TfLiteTensorsToDetectionsCalculator);
 ::mediapipe::Status TfLiteTensorsToDetectionsCalculator::ProcessCPU(
     CalculatorContext* cc, std::vector<Detection>* output_detections) {
   const auto& input_tensors =
-      cc->Inputs().Tag("TENSORS").Get<std::vector<TfLiteTensor>>();
+      cc->Inputs().Tag(kTensorsTag).Get<std::vector<TfLiteTensor>>();
 
   if (input_tensors.size() == 2 ||
       input_tensors.size() == kNumInputTensorsWithAnchors) {
@@ -402,7 +403,7 @@ REGISTER_CALCULATOR(TfLiteTensorsToDetectionsCalculator);
     CalculatorContext* cc, std::vector<Detection>* output_detections) {
 #if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
   const auto& input_tensors =
-      cc->Inputs().Tag("TENSORS_GPU").Get<std::vector<GpuTensor>>();
+      cc->Inputs().Tag(kTensorsGpuTag).Get<std::vector<GpuTensor>>();
   RET_CHECK_GE(input_tensors.size(), 2);
 
   MP_RETURN_IF_ERROR(gpu_helper_.RunInGlContext([this, &input_tensors, &cc,
@@ -466,7 +467,7 @@ REGISTER_CALCULATOR(TfLiteTensorsToDetectionsCalculator);
 #elif defined(MEDIAPIPE_IOS)
 
   const auto& input_tensors =
-      cc->Inputs().Tag("TENSORS_GPU").Get<std::vector<GpuTensor>>();
+      cc->Inputs().Tag(kTensorsGpuTag).Get<std::vector<GpuTensor>>();
   RET_CHECK_GE(input_tensors.size(), 2);
 
   // Copy inputs.

mediapipe/calculators/tflite/tflite_tensors_to_segmentation_calculator.cc

@@ -49,6 +49,16 @@ int NumGroups(const int size, const int group_size) {  // NOLINT
 float Clamp(float val, float min, float max) {
   return std::min(std::max(val, min), max);
 }
+
+constexpr char kTensorsTag[] = "TENSORS";
+constexpr char kTensorsGpuTag[] = "TENSORS_GPU";
+constexpr char kSizeImageTag[] = "REFERENCE_IMAGE";
+constexpr char kSizeImageGpuTag[] = "REFERENCE_IMAGE_GPU";
+constexpr char kMaskTag[] = "MASK";
+constexpr char kMaskGpuTag[] = "MASK_GPU";
+constexpr char kPrevMaskTag[] = "PREV_MASK";
+constexpr char kPrevMaskGpuTag[] = "PREV_MASK_GPU";
+
 }  // namespace
 
 namespace mediapipe {
@@ -148,39 +158,39 @@ REGISTER_CALCULATOR(TfLiteTensorsToSegmentationCalculator);
   bool use_gpu = false;
 
   // Inputs CPU.
-  if (cc->Inputs().HasTag("TENSORS")) {
-    cc->Inputs().Tag("TENSORS").Set<std::vector<TfLiteTensor>>();
+  if (cc->Inputs().HasTag(kTensorsTag)) {
+    cc->Inputs().Tag(kTensorsTag).Set<std::vector<TfLiteTensor>>();
   }
-  if (cc->Inputs().HasTag("PREV_MASK")) {
-    cc->Inputs().Tag("PREV_MASK").Set<ImageFrame>();
+  if (cc->Inputs().HasTag(kPrevMaskTag)) {
+    cc->Inputs().Tag(kPrevMaskTag).Set<ImageFrame>();
   }
-  if (cc->Inputs().HasTag("REFERENCE_IMAGE")) {
-    cc->Inputs().Tag("REFERENCE_IMAGE").Set<ImageFrame>();
+  if (cc->Inputs().HasTag(kSizeImageTag)) {
+    cc->Inputs().Tag(kSizeImageTag).Set<ImageFrame>();
   }
 
   // Inputs GPU.
 #if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
-  if (cc->Inputs().HasTag("TENSORS_GPU")) {
-    cc->Inputs().Tag("TENSORS_GPU").Set<std::vector<GlBuffer>>();
+  if (cc->Inputs().HasTag(kTensorsGpuTag)) {
+    cc->Inputs().Tag(kTensorsGpuTag).Set<std::vector<GlBuffer>>();
     use_gpu |= true;
   }
-  if (cc->Inputs().HasTag("PREV_MASK_GPU")) {
-    cc->Inputs().Tag("PREV_MASK_GPU").Set<mediapipe::GpuBuffer>();
+  if (cc->Inputs().HasTag(kPrevMaskGpuTag)) {
+    cc->Inputs().Tag(kPrevMaskGpuTag).Set<mediapipe::GpuBuffer>();
     use_gpu |= true;
   }
-  if (cc->Inputs().HasTag("REFERENCE_IMAGE_GPU")) {
-    cc->Inputs().Tag("REFERENCE_IMAGE_GPU").Set<mediapipe::GpuBuffer>();
+  if (cc->Inputs().HasTag(kSizeImageGpuTag)) {
+    cc->Inputs().Tag(kSizeImageGpuTag).Set<mediapipe::GpuBuffer>();
     use_gpu |= true;
   }
 #endif  //  !MEDIAPIPE_DISABLE_GPU
 
   // Outputs.
-  if (cc->Outputs().HasTag("MASK")) {
-    cc->Outputs().Tag("MASK").Set<ImageFrame>();
+  if (cc->Outputs().HasTag(kMaskTag)) {
+    cc->Outputs().Tag(kMaskTag).Set<ImageFrame>();
   }
 #if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
-  if (cc->Outputs().HasTag("MASK_GPU")) {
-    cc->Outputs().Tag("MASK_GPU").Set<mediapipe::GpuBuffer>();
+  if (cc->Outputs().HasTag(kMaskGpuTag)) {
+    cc->Outputs().Tag(kMaskGpuTag).Set<mediapipe::GpuBuffer>();
     use_gpu |= true;
   }
 #endif  //  !MEDIAPIPE_DISABLE_GPU
@@ -197,7 +207,7 @@ REGISTER_CALCULATOR(TfLiteTensorsToSegmentationCalculator);
     CalculatorContext* cc) {
   cc->SetOffset(TimestampDiff(0));
 
-  if (cc->Inputs().HasTag("TENSORS_GPU")) {
+  if (cc->Inputs().HasTag(kTensorsGpuTag)) {
     use_gpu_ = true;
 #if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
     MP_RETURN_IF_ERROR(gpu_helper_.Open(cc));
@@ -255,23 +265,22 @@ REGISTER_CALCULATOR(TfLiteTensorsToSegmentationCalculator);
 
 ::mediapipe::Status TfLiteTensorsToSegmentationCalculator::ProcessCpu(
     CalculatorContext* cc) {
-  if (cc->Inputs().Tag("TENSORS").IsEmpty()) {
+  if (cc->Inputs().Tag(kTensorsTag).IsEmpty()) {
     return ::mediapipe::OkStatus();
   }
 
   // Get input streams.
   const auto& input_tensors =
-      cc->Inputs().Tag("TENSORS").Get<std::vector<TfLiteTensor>>();
-  const bool has_prev_mask = cc->Inputs().HasTag("PREV_MASK") &&
-                             !cc->Inputs().Tag("PREV_MASK").IsEmpty();
+      cc->Inputs().Tag(kTensorsTag).Get<std::vector<TfLiteTensor>>();
+  const bool has_prev_mask = cc->Inputs().HasTag(kPrevMaskTag) &&
+                             !cc->Inputs().Tag(kPrevMaskTag).IsEmpty();
   const ImageFrame placeholder;
-  const auto& input_mask = has_prev_mask
-                               ? cc->Inputs().Tag("PREV_MASK").Get<ImageFrame>()
-                               : placeholder;
+  const auto& input_mask =
+      has_prev_mask ? cc->Inputs().Tag(kPrevMaskTag).Get<ImageFrame>()
+                    : placeholder;
   int output_width = tensor_width_, output_height = tensor_height_;
-  if (cc->Inputs().HasTag("REFERENCE_IMAGE")) {
-    const auto& input_image =
-        cc->Inputs().Tag("REFERENCE_IMAGE").Get<ImageFrame>();
+  if (cc->Inputs().HasTag(kSizeImageTag)) {
+    const auto& input_image = cc->Inputs().Tag(kSizeImageTag).Get<ImageFrame>();
     output_width = input_image.Width();
     output_height = input_image.Height();
   }
@@ -353,7 +362,7 @@ REGISTER_CALCULATOR(TfLiteTensorsToSegmentationCalculator);
       ImageFormat::SRGBA, output_width, output_height);
   cv::Mat output_mat = formats::MatView(output_mask.get());
   large_mask_mat.copyTo(output_mat);
-  cc->Outputs().Tag("MASK").Add(output_mask.release(), cc->InputTimestamp());
+  cc->Outputs().Tag(kMaskTag).Add(output_mask.release(), cc->InputTimestamp());
 
   return ::mediapipe::OkStatus();
 }
@@ -364,23 +373,23 @@ REGISTER_CALCULATOR(TfLiteTensorsToSegmentationCalculator);
 // 3. upsample small mask into output mask to be same size as input image
 ::mediapipe::Status TfLiteTensorsToSegmentationCalculator::ProcessGpu(
     CalculatorContext* cc) {
-  if (cc->Inputs().Tag("TENSORS_GPU").IsEmpty()) {
+  if (cc->Inputs().Tag(kTensorsGpuTag).IsEmpty()) {
     return ::mediapipe::OkStatus();
   }
 #if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
   // Get input streams.
   const auto& input_tensors =
-      cc->Inputs().Tag("TENSORS_GPU").Get<std::vector<GlBuffer>>();
-  const bool has_prev_mask = cc->Inputs().HasTag("PREV_MASK_GPU") &&
-                             !cc->Inputs().Tag("PREV_MASK_GPU").IsEmpty();
+      cc->Inputs().Tag(kTensorsGpuTag).Get<std::vector<GlBuffer>>();
+  const bool has_prev_mask = cc->Inputs().HasTag(kPrevMaskGpuTag) &&
+                             !cc->Inputs().Tag(kPrevMaskGpuTag).IsEmpty();
   const auto& input_mask =
       has_prev_mask
-          ? cc->Inputs().Tag("PREV_MASK_GPU").Get<mediapipe::GpuBuffer>()
+          ? cc->Inputs().Tag(kPrevMaskGpuTag).Get<mediapipe::GpuBuffer>()
           : mediapipe::GpuBuffer();
   int output_width = tensor_width_, output_height = tensor_height_;
-  if (cc->Inputs().HasTag("REFERENCE_IMAGE_GPU")) {
+  if (cc->Inputs().HasTag(kSizeImageGpuTag)) {
     const auto& input_image =
-        cc->Inputs().Tag("REFERENCE_IMAGE_GPU").Get<mediapipe::GpuBuffer>();
+        cc->Inputs().Tag(kSizeImageGpuTag).Get<mediapipe::GpuBuffer>();
     output_width = input_image.width();
     output_height = input_image.height();
   }
@@ -441,7 +450,7 @@ REGISTER_CALCULATOR(TfLiteTensorsToSegmentationCalculator);
   // Send out image as GPU packet.
   auto output_image = output_texture.GetFrame<mediapipe::GpuBuffer>();
   cc->Outputs()
-      .Tag("MASK_GPU")
+      .Tag(kMaskGpuTag)
       .Add(output_image.release(), cc->InputTimestamp());
 
   // Cleanup