Add a preliminary global parallelizer for heavy inference. Parallelize the silhouette matcher.

Author: Mysticial

Common/Cpp/Concurrency/ParallelTaskRunner.cpp

@@ -10,6 +10,10 @@
 #include "Common/Cpp/PanicDump.h"
 #include "ParallelTaskRunner.h"
 
+//#include <iostream>
+//using std::cout;
+//using std::endl;
+
 namespace PokemonAutomation{
 
 
@@ -72,6 +76,37 @@ std::shared_ptr<AsyncTask> ParallelTaskRunner::dispatch(std::function<void()>&&
 }
 
 
+void ParallelTaskRunner::run_in_parallel(
+    const std::function<void(size_t index)>& func,
+    size_t start, size_t end,
+    size_t block_size
+){
+    if (start >= end){
+        return;
+    }
+    size_t total = end - start;
+    size_t blocks = (total + block_size - 1) / block_size;
+
+    std::vector<std::shared_ptr<AsyncTask>> tasks;
+    for (size_t c = 0; c < blocks; c++){
+        tasks.emplace_back(dispatch([=, &func]{
+            size_t s = start + c * block_size;
+            size_t e = std::min(s + block_size, end);
+//            cout << "Running: [" << s << "," << e << ")" << endl;
+            for (; s < e; s++){
+                func(s);
+            }
+        }));
+    }
+
+    for (std::shared_ptr<AsyncTask>& task : tasks){
+        task->wait_and_rethrow_exceptions();
+    }
+}
+
+
+
+
 void ParallelTaskRunner::thread_loop(){
 //#if _WIN32
 //    SetThreadPriority(GetCurrentThread(), THREAD_PRIORITY_IDLE);

Common/Cpp/Concurrency/ParallelTaskRunner.h

@@ -25,6 +25,12 @@ class ParallelTaskRunner{
 
     std::shared_ptr<AsyncTask> dispatch(std::function<void()>&& func);
 
+    void run_in_parallel(
+        const std::function<void(size_t index)>& func,
+        size_t start, size_t end,
+        size_t block_size = 0
+    );
+
 
 private:
 //    void dispatch_task(AsyncTask& task);

SerialPrograms/CMakeLists.txt

@@ -173,9 +173,7 @@ file(GLOB MAIN_SOURCES
     ../Common/Microcontroller/DeviceRoutines.h
     ../Common/Microcontroller/MessageProtocol.h
     ../Common/NintendoSwitch/NintendoSwitch_ControllerDefs.h
-    ../Common/NintendoSwitch/NintendoSwitch_Protocol_DigitEntry.h
     ../Common/NintendoSwitch/NintendoSwitch_Protocol_PushButtons.h
-    ../Common/NintendoSwitch/NintendoSwitch_Protocol_Routines.h
     ../Common/NintendoSwitch/NintendoSwitch_Protocol_Superscalar.h
     ../Common/NintendoSwitch/NintendoSwitch_SlotDatabase.h
     ../Common/PokemonSwSh/PokemonProgramIDs.h
@@ -535,6 +533,8 @@ file(GLOB MAIN_SOURCES
     Source/CommonTools/InferencePivots/AudioInferencePivot.h
     Source/CommonTools/InferencePivots/VisualInferencePivot.cpp
     Source/CommonTools/InferencePivots/VisualInferencePivot.h
+    Source/CommonTools/GlobalInferenceRunner.cpp
+    Source/CommonTools/GlobalInferenceRunner.h
     Source/CommonTools/InferenceThrottler.h
     Source/CommonTools/OCR/OCR_DictionaryMatcher.cpp
     Source/CommonTools/OCR/OCR_DictionaryMatcher.h

SerialPrograms/SerialPrograms.pro

@@ -254,6 +254,7 @@ SOURCES += \
     Source/CommonTools/Audio/AudioPerSpectrumDetectorBase.cpp \
     Source/CommonTools/Audio/AudioTemplateCache.cpp \
     Source/CommonTools/Audio/SpectrogramMatcher.cpp \
+    Source/CommonTools/GlobalInferenceRunner.cpp \
     Source/CommonTools/ImageMatch/CroppedImageDictionaryMatcher.cpp \
     Source/CommonTools/ImageMatch/ExactImageDictionaryMatcher.cpp \
     Source/CommonTools/ImageMatch/ExactImageMatcher.cpp \
@@ -1184,9 +1185,7 @@ HEADERS += \
     ../Common/Microcontroller/DeviceRoutines.h \
     ../Common/Microcontroller/MessageProtocol.h \
     ../Common/NintendoSwitch/NintendoSwitch_ControllerDefs.h \
-    ../Common/NintendoSwitch/NintendoSwitch_Protocol_DigitEntry.h \
     ../Common/NintendoSwitch/NintendoSwitch_Protocol_PushButtons.h \
-    ../Common/NintendoSwitch/NintendoSwitch_Protocol_Routines.h \
     ../Common/NintendoSwitch/NintendoSwitch_Protocol_Superscalar.h \
     ../Common/NintendoSwitch/NintendoSwitch_SlotDatabase.h \
     ../Common/PokemonSwSh/PokemonProgramIDs.h \
@@ -1362,6 +1361,7 @@ HEADERS += \
     Source/CommonTools/Audio/AudioTemplateCache.h \
     Source/CommonTools/Audio/SpectrogramMatcher.h \
     Source/CommonTools/DetectionDebouncer.h \
+    Source/CommonTools/GlobalInferenceRunner.h \
     Source/CommonTools/ImageMatch/CroppedImageDictionaryMatcher.h \
     Source/CommonTools/ImageMatch/ExactImageDictionaryMatcher.h \
     Source/CommonTools/ImageMatch/ExactImageMatcher.h \

SerialPrograms/Source/CommonTools/GlobalInferenceRunner.cpp

@@ -0,0 +1,27 @@
+/*  Global Inference Runner
+ *
+ *  From: https://github.com/PokemonAutomation/Arduino-Source
+ *
+ */
+
+#include "CommonFramework/GlobalSettingsPanel.h"
+#include "CommonFramework/Options/Environment/PerformanceOptions.h"
+#include "GlobalInferenceRunner.h"
+
+namespace PokemonAutomation{
+
+
+
+ParallelTaskRunner& global_inference_runner(){
+    static ParallelTaskRunner runner(
+        [](){
+            GlobalSettings::instance().PERFORMANCE->INFERENCE_PRIORITY.set_on_this_thread();
+        },
+        0, std::thread::hardware_concurrency()
+    );
+    return runner;
+}
+
+
+
+}

SerialPrograms/Source/CommonTools/GlobalInferenceRunner.h

@@ -0,0 +1,21 @@
+/*  Global Inference Runner
+ *
+ *  From: https://github.com/PokemonAutomation/Arduino-Source
+ *
+ */
+
+#ifndef PokemonAutomation_CommonTools_GlobalInferenceRunner_H
+#define PokemonAutomation_CommonTools_GlobalInferenceRunner_H
+
+#include "Common/Cpp/Concurrency/ParallelTaskRunner.h"
+
+namespace PokemonAutomation{
+
+
+
+ParallelTaskRunner& global_inference_runner();
+
+
+
+}
+#endif

SerialPrograms/Source/CommonTools/ImageMatch/SilhouetteDictionaryMatcher.cpp

@@ -5,8 +5,10 @@
  */
 
 #include "Common/Cpp/Exceptions.h"
+#include "Common/Cpp/Concurrency/SpinLock.h"
 #include "CommonFramework/ImageTypes/ImageViewRGB32.h"
 #include "CommonFramework/ImageTools/ImageDiff.h"
+#include "CommonTools/GlobalInferenceRunner.h"
 #include "ImageCropper.h"
 #include "SilhouetteDictionaryMatcher.h"
 
@@ -27,11 +29,12 @@ void SilhouetteDictionaryMatcher::add(const std::string& slug, const ImageViewRG
         throw InternalProgramError(nullptr, PA_CURRENT_FUNCTION, "Duplicate slug: " + slug);
     }
 
-    m_database.emplace(
+    iter = m_database.emplace(
         std::piecewise_construct,
         std::forward_as_tuple(slug),
         std::forward_as_tuple(trim_image_alpha(image).copy())
-    );
+    ).first;
+    m_database_vector.emplace_back(&*iter);
 }
 
 
@@ -45,14 +48,31 @@ ImageMatchResult SilhouetteDictionaryMatcher::match(
         return results;
     }
 
+    SpinLock lock;
+    global_inference_runner().run_in_parallel(
+        [&](size_t index){
+            const auto& matcher = *m_database_vector[index];
+            double alpha = matcher.second.rmsd_masked(image);
+            WriteSpinLock lg(lock);
+            results.add(alpha, matcher.first);
+            results.clear_beyond_spread(alpha_spread);
+        },
+        0, m_database_vector.size(),
+        100
+    );
+
+
+#if 0
     for (const auto& item : m_database){
 //        if (item.first != "solosis"){
 //            continue;
 //        }
         double alpha = item.second.rmsd_masked(image);
+//        WriteSpinLock lg(lock);
         results.add(alpha, item.first);
         results.clear_beyond_spread(alpha_spread);
     }
+#endif
 
     return results;
 }

SerialPrograms/Source/CommonTools/ImageMatch/SilhouetteDictionaryMatcher.h

@@ -47,6 +47,7 @@ class SilhouetteDictionaryMatcher{
 
 private:
     std::map<std::string, ExactImageMatcher> m_database;
+    std::vector<const std::pair<const std::string, ExactImageMatcher>*> m_database_vector;
 };