バッチサイズ1の場合並列化しない

cppshogi/python_module.cpp

@@ -32,12 +32,12 @@ inline T is_nyugyoku(const uint8_t result) {
 }
 
 void __hcpe_decode_with_value(const size_t len, char* ndhcpe, char* ndfeatures1, char* ndfeatures2, char* ndmove, char* ndresult, char* ndvalue) {
-	HuffmanCodedPosAndEval *hcpe = reinterpret_cast<HuffmanCodedPosAndEval *>(ndhcpe);
+	HuffmanCodedPosAndEval* hcpe = reinterpret_cast<HuffmanCodedPosAndEval*>(ndhcpe);
 	features1_t* features1 = reinterpret_cast<features1_t*>(ndfeatures1);
 	features2_t* features2 = reinterpret_cast<features2_t*>(ndfeatures2);
 	int64_t* move = reinterpret_cast<int64_t*>(ndmove);
-	float *result = reinterpret_cast<float *>(ndresult);
-	float *value = reinterpret_cast<float *>(ndvalue);
+	float* result = reinterpret_cast<float*>(ndresult);
+	float* value = reinterpret_cast<float*>(ndvalue);
 
 	// set all zero
 	std::fill_n((float*)features1, sizeof(features1_t) / sizeof(float) * len, 0.0f);
@@ -62,7 +62,7 @@ void __hcpe_decode_with_value(const size_t len, char* ndhcpe, char* ndfeatures1,
 }
 
 void __hcpe2_decode_with_value(const size_t len, char* ndhcpe2, char* ndfeatures1, char* ndfeatures2, char* ndmove, char* ndresult, char* ndvalue, char* ndaux) {
-	HuffmanCodedPosAndEval2 *hcpe = reinterpret_cast<HuffmanCodedPosAndEval2 *>(ndhcpe2);
+	HuffmanCodedPosAndEval2* hcpe = reinterpret_cast<HuffmanCodedPosAndEval2*>(ndhcpe2);
 	features1_t* features1 = reinterpret_cast<features1_t*>(ndfeatures1);
 	features2_t* features2 = reinterpret_cast<features2_t*>(ndfeatures2);
 	int64_t* move = reinterpret_cast<int64_t*>(ndmove);
@@ -144,17 +144,16 @@ void __hcpe3_create_cache(const std::string& filepath) {
 }
 
 // hcpe3キャッシュ
-std::ifstream cache;
+std::ifstream* cache;
 std::vector<size_t> cache_pos;
-std::mutex cache_mutex;
 size_t __hcpe3_load_cache(const std::string& filepath) {
-	cache.open(filepath, std::ios::binary);
+	cache = new std::ifstream(filepath, std::ios::binary);
 	size_t num;
-	cache.read((char*)&num, sizeof(num));
+	cache->read((char*)&num, sizeof(num));
 	cache_pos.resize(num + 1);
-	cache.read((char*)cache_pos.data(), sizeof(size_t) * num);
-	cache.seekg(0, std::ios_base::end);
-	cache_pos[num] = cache.tellg();
+	cache->read((char*)cache_pos.data(), sizeof(size_t) * num);
+	cache->seekg(0, std::ios_base::end);
+	cache_pos[num] = cache->tellg();
 	return num;
 }
 
@@ -165,14 +164,27 @@ size_t __hcpe3_get_cache_num() {
 TrainingData get_cache(const size_t i) {
 	const size_t pos = cache_pos[i];
 	const size_t candidateNum = ((cache_pos[i + 1] - pos) - sizeof(Hcpe3CacheBody)) / sizeof(Hcpe3CacheCandidate);
-	cache_mutex.lock();
-	cache.seekg(pos, std::ios_base::beg);
 	struct Hcpe3CacheBuf {
 		Hcpe3CacheBody body;
 		Hcpe3CacheCandidate candidates[MaxLegalMoves];
 	} buf;
-	cache.read((char*)&buf, sizeof(Hcpe3CacheBody) + sizeof(Hcpe3CacheCandidate) * candidateNum);
-	cache_mutex.unlock();
+	cache->seekg(pos, std::ios_base::beg);
+	cache->read((char*)&buf, sizeof(Hcpe3CacheBody) + sizeof(Hcpe3CacheCandidate) * candidateNum);
+	return TrainingData(buf.body, buf.candidates, candidateNum);
+}
+
+TrainingData get_cache_with_lock(const size_t i) {
+	const size_t pos = cache_pos[i];
+	const size_t candidateNum = ((cache_pos[i + 1] - pos) - sizeof(Hcpe3CacheBody)) / sizeof(Hcpe3CacheCandidate);
+	struct Hcpe3CacheBuf {
+		Hcpe3CacheBody body;
+		Hcpe3CacheCandidate candidates[MaxLegalMoves];
+	} buf;
+	#pragma omp critical
+	{
+		cache->seekg(pos, std::ios_base::beg);
+		cache->read((char*)&buf, sizeof(Hcpe3CacheBody) + sizeof(Hcpe3CacheCandidate) * candidateNum);
+	}
 	return TrainingData(buf.body, buf.candidates, candidateNum);
 }
 
@@ -418,9 +430,9 @@ void __hcpe3_decode_with_value(const size_t len, char* ndindex, char* ndfeatures
 	std::fill_n((float*)features2, sizeof(features2_t) / sizeof(float) * len, 0.0f);
 	std::fill_n((float*)probability, 9 * 9 * MAX_MOVE_LABEL_NUM * len, 0.0f);
 
-	#pragma omp parallel for num_threads(2)
-	for (int64_t i = 0; i < len; i++) {
-		const auto& hcpe3 = cache.is_open() ? get_cache(index[i]) : trainingData[index[i]];
+	#pragma omp parallel for num_threads(2) if (len > 1)
+	for (size_t i = 0; i < len; i++) {
+		const auto& hcpe3 = cache ? (len > 1 ? get_cache_with_lock(index[i]) : get_cache(index[i])) : trainingData[index[i]];
 
 		Position position;
 		position.set(hcpe3.hcp);
@@ -447,10 +459,10 @@ void __hcpe3_decode_with_value(const size_t len, char* ndindex, char* ndfeatures
 void __hcpe3_get_hcpe(const size_t index, char* ndhcpe) {
 	HuffmanCodedPosAndEval* hcpe = reinterpret_cast<HuffmanCodedPosAndEval*>(ndhcpe);
 
-	const auto& hcpe3 = cache.is_open() ? get_cache(index) : trainingData[index];
+	const auto& hcpe3 = cache ? get_cache(index) : trainingData[index];
 
 	hcpe->hcp = hcpe3.hcp;
-	float max_prob = FLT_MIN ;
+	float max_prob = FLT_MIN;
 	for (const auto kv : hcpe3.candidates) {
 		const auto& move16 = kv.first;
 		const auto& prob = kv.second;