Project import generated by Copybara.

GitOrigin-RevId: ec25bf2e416c3689477e82946fb69de2e53b9161

mediapipe/calculators/core/end_loop_calculator.cc

@@ -28,6 +28,10 @@ typedef EndLoopCalculator<std::vector<::mediapipe::NormalizedRect>>
     EndLoopNormalizedRectCalculator;
 REGISTER_CALCULATOR(EndLoopNormalizedRectCalculator);
 
+typedef EndLoopCalculator<std::vector<::mediapipe::LandmarkList>>
+    EndLoopLandmarkListVectorCalculator;
+REGISTER_CALCULATOR(EndLoopLandmarkListVectorCalculator);
+
 typedef EndLoopCalculator<std::vector<::mediapipe::NormalizedLandmarkList>>
     EndLoopNormalizedLandmarkListVectorCalculator;
 REGISTER_CALCULATOR(EndLoopNormalizedLandmarkListVectorCalculator);

mediapipe/calculators/tensor/inference_calculator_cpu.cc

@@ -35,20 +35,28 @@ namespace api2 {
 
 namespace {
 
+int GetXnnpackDefaultNumThreads() {
+#if defined(MEDIAPIPE_ANDROID) || defined(MEDIAPIPE_IOS) || \
+    defined(__EMSCRIPTEN_PTHREADS__)
+  constexpr int kMinNumThreadsByDefault = 1;
+  constexpr int kMaxNumThreadsByDefault = 4;
+  return std::clamp(NumCPUCores() / 2, kMinNumThreadsByDefault,
+                    kMaxNumThreadsByDefault);
+#else
+  return 1;
+#endif  // MEDIAPIPE_ANDROID || MEDIAPIPE_IOS || __EMSCRIPTEN_PTHREADS__
+}
+
 // Returns number of threads to configure XNNPACK delegate with.
-// (Equal to user provided value if specified.  Otherwise, it returns number of
-// high cores (hard-coded to 1 for Emscripten without Threads extension))
+// Returns user provided value if specified. Otherwise, tries to choose optimal
+// number of threads depending on the device.
 int GetXnnpackNumThreads(const mediapipe::InferenceCalculatorOptions& opts) {
   static constexpr int kDefaultNumThreads = -1;
   if (opts.has_delegate() && opts.delegate().has_xnnpack() &&
       opts.delegate().xnnpack().num_threads() != kDefaultNumThreads) {
     return opts.delegate().xnnpack().num_threads();
   }
-#if !defined(__EMSCRIPTEN__) || defined(__EMSCRIPTEN_PTHREADS__)
-  return InferHigherCoreIds().size();
-#else
-  return 1;
-#endif  // !__EMSCRIPTEN__ || __EMSCRIPTEN_PTHREADS__
+  return GetXnnpackDefaultNumThreads();
 }
 
 }  // namespace

mediapipe/calculators/tensor/inference_calculator_gl.cc

@@ -269,8 +269,8 @@ absl::Status InferenceCalculatorGlImpl::InitTFLiteGPURunner(
       break;
     }
   }
-  MP_RETURN_IF_ERROR(
-      tflite_gpu_runner_->InitializeWithModel(model, op_resolver));
+  MP_RETURN_IF_ERROR(tflite_gpu_runner_->InitializeWithModel(
+      model, op_resolver, /*allow_quant_ops=*/true));
 
   // Create and bind OpenGL buffers for outputs.
   // The buffers are created once and their ids are passed to calculator outputs

mediapipe/calculators/tensor/inference_calculator_metal.cc

@@ -226,6 +226,10 @@ absl::Status InferenceCalculatorMetalImpl::LoadDelegate(CalculatorContext* cc) {
 
   // Configure and create the delegate.
   TFLGpuDelegateOptions options;
+  // `enable_quantization` enables the run of sparse models i.e. the models with
+  // DENSIFY op preceding DEQUINTIZE op. Both ops get removed from the execution
+  // graph after the tensor of the weights is read.
+  options.enable_quantization = true;
   options.allow_precision_loss = allow_precision_loss_;
   options.wait_type = TFLGpuDelegateWaitType::TFLGpuDelegateWaitTypeDoNotWait;
   delegate_ =

mediapipe/calculators/tensor/tensors_to_segmentation_calculator.cc

@@ -763,9 +763,13 @@ out vec4 fragColor;
 #endif  // defined(GL_ES);
 
 void main() {
-
-  vec4 input_value = texture2D(input_texture, sample_coordinate);
-  vec2 gid = sample_coordinate;
+#ifdef FLIP_Y_COORD
+  float y_coord = 1.0 - sample_coordinate.y;
+#else
+  float y_coord = sample_coordinate.y;
+#endif  // defined(FLIP_Y_COORD)
+  vec2 adjusted_coordinate = vec2(sample_coordinate.x, y_coord);
+  vec4 input_value = texture2D(input_texture, adjusted_coordinate);
 
   // Run activation function.
   // One and only one of FN_SOFTMAX,FN_SIGMOID,FN_NONE will be defined.
@@ -787,13 +791,6 @@ void main() {
   float new_mask_value = input_value.r;
 #endif // FN_NONE
 
-#ifdef FLIP_Y_COORD
-  float y_coord = 1.0 - gid.y;
-#else
-  float y_coord = gid.y;
-#endif  // defined(FLIP_Y_COORD)
-  vec2 output_coordinate = vec2(gid.x, y_coord);
-
   vec4 out_value = vec4(new_mask_value, 0.0, 0.0, new_mask_value);
   fragColor = out_value;
 })";

mediapipe/calculators/tflite/tflite_inference_calculator.cc

@@ -128,9 +128,23 @@ struct GPUData {
 }  // namespace
 #endif  // MEDIAPIPE_TFLITE_GPU_SUPPORTED
 
+namespace {
+
+int GetXnnpackDefaultNumThreads() {
+#if defined(MEDIAPIPE_ANDROID) || defined(MEDIAPIPE_IOS) || \
+    defined(__EMSCRIPTEN_PTHREADS__)
+  constexpr int kMinNumThreadsByDefault = 1;
+  constexpr int kMaxNumThreadsByDefault = 4;
+  return std::clamp(NumCPUCores() / 2, kMinNumThreadsByDefault,
+                    kMaxNumThreadsByDefault);
+#else
+  return 1;
+#endif  // MEDIAPIPE_ANDROID || MEDIAPIPE_IOS || __EMSCRIPTEN_PTHREADS__
+}
+
 // Returns number of threads to configure XNNPACK delegate with.
-// (Equal to user provided value if specified.  Otherwise, it returns number of
-// high cores (hard-coded to 1 for Emscripten without Threads extension))
+// Returns user provided value if specified. Otherwise, tries to choose optimal
+// number of threads depending on the device.
 int GetXnnpackNumThreads(
     const mediapipe::TfLiteInferenceCalculatorOptions& opts) {
   static constexpr int kDefaultNumThreads = -1;
@@ -138,13 +152,11 @@ int GetXnnpackNumThreads(
       opts.delegate().xnnpack().num_threads() != kDefaultNumThreads) {
     return opts.delegate().xnnpack().num_threads();
   }
-#if !defined(__EMSCRIPTEN__) || defined(__EMSCRIPTEN_PTHREADS__)
-  return InferHigherCoreIds().size();
-#else
-  return 1;
-#endif  // !__EMSCRIPTEN__ || __EMSCRIPTEN_PTHREADS__
+  return GetXnnpackDefaultNumThreads();
 }
 
+}  // namespace
+
 // Calculator Header Section
 
 // Runs inference on the provided input TFLite tensors and TFLite model.
@@ -737,8 +749,8 @@ absl::Status TfLiteInferenceCalculator::InitTFLiteGPURunner(
       break;
     }
   }
-  MP_RETURN_IF_ERROR(
-      tflite_gpu_runner_->InitializeWithModel(model, *op_resolver_ptr));
+  MP_RETURN_IF_ERROR(tflite_gpu_runner_->InitializeWithModel(
+      model, *op_resolver_ptr, /*allow_quant_ops=*/true));
 
   // Allocate interpreter memory for cpu output.
   if (!gpu_output_) {
@@ -969,6 +981,10 @@ absl::Status TfLiteInferenceCalculator::LoadDelegate(CalculatorContext* cc) {
   const int kHalfSize = 2;  // sizeof(half)
   // Configure and create the delegate.
   TFLGpuDelegateOptions options;
+  // `enable_quantization` enables the run of sparse models i.e. the models with
+  // DENSIFY op preceding DEQUINTIZE op. Both ops get removed from the execution
+  // graph after the tensor of the weights is read.
+  options.enable_quantization = true;
   options.allow_precision_loss = allow_precision_loss_;
   options.wait_type = TFLGpuDelegateWaitType::TFLGpuDelegateWaitTypeActive;
   if (!delegate_)

mediapipe/calculators/util/filter_collection_calculator.cc

@@ -32,11 +32,15 @@ typedef FilterCollectionCalculator<std::vector<::mediapipe::NormalizedRect>>
     FilterNormalizedRectCollectionCalculator;
 REGISTER_CALCULATOR(FilterNormalizedRectCollectionCalculator);
 
-typedef FilterCollectionCalculator<
-    std::vector<::mediapipe::NormalizedLandmarkList>>
+typedef FilterCollectionCalculator<std::vector<::mediapipe::LandmarkList>>
     FilterLandmarkListCollectionCalculator;
 REGISTER_CALCULATOR(FilterLandmarkListCollectionCalculator);
 
+typedef FilterCollectionCalculator<
+    std::vector<::mediapipe::NormalizedLandmarkList>>
+    FilterNormalizedLandmarkListCollectionCalculator;
+REGISTER_CALCULATOR(FilterNormalizedLandmarkListCollectionCalculator);
+
 typedef FilterCollectionCalculator<std::vector<::mediapipe::ClassificationList>>
     FilterClassificationListCollectionCalculator;
 REGISTER_CALCULATOR(FilterClassificationListCollectionCalculator);