reduce memory usage on score function, longlong => char

qubic · Apr 12, 2024 · 62a1153 · 62a1153
1 parent 2f98f05
commit 62a1153
Show file tree

Hide file tree

Showing 3 changed files with 26 additions and 17 deletions.
diff --git a/src/score.h b/src/score.h
@@ -24,11 +24,12 @@ template<
 struct ScoreFunction
 {
     long long miningData[dataLength];
+    //neuron only has values [-1, 0, 1]
     struct
     {
-        long long input[dataLength + numberOfInputNeurons + infoLength];
-        long long output[infoLength + numberOfOutputNeurons + dataLength];
-        long long buffer[dataLength + numberOfInputNeurons + infoLength];
+        char input[dataLength + numberOfInputNeurons + infoLength];
+        char output[infoLength + numberOfOutputNeurons + dataLength];
+        char buffer[dataLength + numberOfInputNeurons + infoLength];
     } _neurons[solutionBufferCount];
     struct
     {
@@ -52,7 +53,7 @@ struct ScoreFunction
     int _bufferPos[solutionBufferCount][numberOfOutputNeurons + dataLength][129];
 #endif
     int nSample;
-    long long _sumBuffer[solutionBufferCount][dataLength + numberOfInputNeurons + infoLength];
+    char _sumBuffer[solutionBufferCount][dataLength + numberOfInputNeurons + infoLength];
     unsigned short _indices[solutionBufferCount][dataLength + numberOfInputNeurons + infoLength];
 
     m256i initialRandomSeed;
@@ -261,7 +262,7 @@ struct ScoreFunction
     }
 
     void getLastNeurons(const unsigned short* indices, const int* bucket, const int* modNum, unsigned short* topMax,
-                        const int nMax, const int totalModNum, int& currentCount, long long* neuron, int* index)
+                        const int nMax, const int totalModNum, int& currentCount, char* neuron, int* index)
     {
         while (currentCount < nMax)
         {
@@ -285,7 +286,7 @@ struct ScoreFunction
     }
 
     short getLastNeuronsIndex(const unsigned short* indices, const int* bucket, const int* modNum,
-                              const int totalModNum, long long* neuron, int* index)
+                              const int totalModNum, char* neuron, int* index)
     {
         int current_max = -1;
         int max_id = -1;
@@ -359,7 +360,7 @@ struct ScoreFunction
     template <bool isInput, int beginLength, int neuronLength, int endLength, int duration>
     void computeNeuron(int solutionBufIdx)
     {
-        long long* neurons = nullptr;
+        char* neurons = nullptr;
         char* synapses = nullptr;
         if (isInput) {
             neurons = _neurons[solutionBufIdx].input;
@@ -462,7 +463,10 @@ struct ScoreFunction
 
         // compute input
         setMem(&neurons.input[0], sizeof(neurons.input), 0);
-        copyMem(&neurons.input[0], miningData, sizeof(miningData));
+        for (int i = 0; i < dataLength; i++)
+        {
+            neurons.input[i] = (char)(miningData[i]);
+        }
         computeBucket<true, dataLength, numberOfInputNeurons, infoLength>(solutionBufIdx);
         computeNeuron<true, dataLength, numberOfInputNeurons, infoLength, maxInputDuration>(solutionBufIdx);
 

diff --git a/test/score.cpp b/test/score.cpp
@@ -76,16 +76,16 @@ struct ScoreTester
         std::cout << "current score() returns " << current << ", reference score() returns " << reference << std::endl;
         bool matchedInputNeuron = true;
         bool matchedOutputNeuron = true;
-        for (int i = 0; i < 1024 + 2048 + 512; i++) {
+        for (int i = 0; i < dataLength + numberOfInputNeurons + infoLength; i++) {
             if ((*score)._neurons[0].input[i] != (*score_ref_impl)._neurons[0].input[i]) {
-                printf("DIFF Input neuron %d: %lld vs %lld\n", i, (*score)._neurons[0].input[i], (*score_ref_impl)._neurons[0].input[i]);
+                printf("DIFF Input neuron %d: %d vs %d\n", i, (*score)._neurons[0].input[i], (*score_ref_impl)._neurons[0].input[i]);
                 matchedInputNeuron = false;
                 break;
             }
         }
-        for (int i = 0; i < 1024 + 2048 + 512; i++) {
+        for (int i = 0; i < dataLength + numberOfOutputNeurons + infoLength; i++) {
             if ((*score)._neurons[0].output[i] != (*score_ref_impl)._neurons[0].output[i]) {
-                printf("DIFF Output neuron %d: %lld vs %lld\n", i, (*score)._neurons[0].output[i], (*score_ref_impl)._neurons[0].output[i]);
+                printf("DIFF Output neuron %d: %d vs %d\n", i, (*score)._neurons[0].output[i], (*score_ref_impl)._neurons[0].output[i]);
                 matchedOutputNeuron = false;
                 break;
             }

diff --git a/test/score_reference.h b/test/score_reference.h
@@ -17,11 +17,13 @@ template<
 struct ScoreReferenceImplementation
 {
     long long miningData[dataLength];
+
+    //neuron only has values [-1, 0, 1]
     struct
     {
-        long long input[dataLength + numberOfInputNeurons + infoLength];
-        long long output[infoLength + numberOfOutputNeurons + dataLength];
-        long long neuronBuffer[dataLength + numberOfInputNeurons + infoLength];
+        char input[dataLength + numberOfInputNeurons + infoLength];
+        char output[infoLength + numberOfOutputNeurons + dataLength];
+        char neuronBuffer[dataLength + numberOfInputNeurons + infoLength];
     } _neurons[solutionBufferCount];
     struct
     {
@@ -41,7 +43,7 @@ struct ScoreReferenceImplementation
         }
     }
 
-    static inline void clampNeuron(long long& val)
+    static inline void clampNeuron(char& val)
     {
         if (val > NEURON_VALUE_LIMIT) {
             val = NEURON_VALUE_LIMIT;
@@ -90,7 +92,10 @@ struct ScoreReferenceImplementation
             synapses.outputLength[outputNeuronIndex * (infoLength + numberOfOutputNeurons + dataLength) + (infoLength + outputNeuronIndex)] = 0;
         }
 
-        memcpy(&neurons.input[0], &miningData, sizeof(miningData));        
+        for (int i = 0; i < dataLength; i++)
+        {
+            neurons.input[i] = (char)(miningData[i]);
+        }
 
         for (int tick = 1; tick <= maxInputDuration; tick++)
         {