Merge pull request #54 from GPUEngineering/feature/53-serialisation-simple

DTensor from text file

Author: alphaville

CHANGELOG.md

@@ -5,6 +5,16 @@ All notable changes to this project will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+<!-- ---------------------
+      v1.5.0
+     --------------------- -->
+## v1.5.0 - 27-11-2024
+
+### Added
+
+- Created methods for serialising and deserialising `DTensor` objects
+
+
 <!-- ---------------------
       v1.4.0
      --------------------- -->

README.md

@@ -246,6 +246,27 @@ The `DTensor` `B` will be overwritten with the solution.
 > overwrite only part of the given `B`, as `B` is a
 > (4,1,1)-tensor and the solution is a (3,1,1)-tensor.
 
+### 1.8. Saving and loading tensors
+
+Tensor data can be stored in simple text files which have the following structure
+
+```text
+number_of_rows
+number_of_columns
+number_of_matrices
+data (one entry per line)
+```
+
+To save a tensor in a file, simply call `DTensor::saveToFile(filename)`.
+
+To load a tensor from a file, the static function `DTensor<T>::parseFromTextFile(filename)` can be used. For example:
+
+```c++
+auto z = DTensor<double>::parseFromTextFile("path/to/my.dtensor")
+```
+
+If necessary, you can provide a second argument to `parseFromTextFile` to specify the order in which the data are stored (the `StorageMode`). 
+
 ## 2. Cholesky factorisation and system solution
 
 > [!WARNING]

include/tensor.cuh

@@ -11,6 +11,7 @@
 #include <memory>
 #include <optional>
 #include <cassert>
+#include <fstream>
 
 #ifndef TENSOR_CUH
 #define TENSOR_CUH
@@ -250,6 +251,19 @@ public:
      */
     static DTensor<T> createRandomTensor(size_t numRows, size_t numCols, size_t numMats, T low, T hi);
 
+    /**
+     * Parse data from text file and create an instance of DTensor
+     *
+     * This static function reads data from a text file, creates a DTensor and uploads the data to the device.
+     *
+     * @param path_to_file path to file as string
+     * @param mode storage mode (default: StorageMode::defaultMajor)
+     * @return instance of DTensor
+     *
+     * @throws std::invalid_argument if the file is not found
+     */
+    static DTensor<T> parseFromTextFile(std::string path_to_file, StorageMode mode = StorageMode::defaultMajor);
+
     /**
     * Constructs a DTensor object.
     */
@@ -487,6 +501,13 @@ public:
      */
     void reshape(size_t newNumRows, size_t newNumCols, size_t newNumMats = 1);
 
+    /**
+     * Saves the current instance of DTensor to a (text) file
+     *
+     * @param pathToFile
+     */
+    void saveToFile(std::string pathToFile);
+
     /* ------------- OPERATORS ------------- */
 
     DTensor &operator=(const DTensor &other);
@@ -564,6 +585,80 @@ DTensor<T> DTensor<T>::createRandomTensor(size_t numRows, size_t numCols, size_t
     throw std::invalid_argument("[createRandomTensor] unsupported type T");
 }
 
+
+template<typename T>
+struct data_t {
+    size_t numRows;
+    size_t numCols;
+    size_t numMats;
+    std::vector<T> data;
+};
+
+template<typename T>
+data_t<T> vectorFromFile(std::string path_to_file) {
+    data_t<T> dataStruct;
+    std::ifstream file;
+    file.open(path_to_file, std::ios::in);
+    if (!file.is_open()) { throw std::invalid_argument("the file you provided does not exist"); };
+
+    std::string line;
+    getline(file, line); dataStruct.numRows = atoi(line.c_str());
+    getline(file, line); dataStruct.numCols = atoi(line.c_str());
+    getline(file, line); dataStruct.numMats = atoi(line.c_str());
+
+    size_t numElements = dataStruct.numRows * dataStruct.numCols * dataStruct.numMats;
+    std::vector<T> vecDataFromFile(numElements);
+
+    size_t i = 0;
+    while (getline(file, line)) {
+        if constexpr (std::is_same_v<T, int>) {
+            vecDataFromFile[i] = atoi(line.c_str());
+        } else if constexpr (std::is_same_v<T, double>) {
+            vecDataFromFile[i] = std::stod(line.c_str());
+        } else if constexpr (std::is_same_v<T, float>) {
+            vecDataFromFile[i] = std::stof(line.c_str());
+        } else if constexpr (std::is_same_v<T, long double>) {
+            vecDataFromFile[i] = std::stold(line.c_str());
+        } else if constexpr (std::is_same_v<T, long>) {
+            vecDataFromFile[i] = std::stol(line.c_str());
+        } else if constexpr (std::is_same_v<T, long long>) {
+            vecDataFromFile[i] = std::stoll(line.c_str());
+        } else if constexpr (std::is_same_v<T, unsigned long>) {
+            vecDataFromFile[i] = std::stoul(line.c_str());
+        } else if constexpr (std::is_same_v<T, unsigned long long>) {
+            vecDataFromFile[i] = std::stoull(line.c_str());
+        } else if constexpr (std::is_same_v<T, size_t>) {
+            sscanf(line.c_str(), "%zu", &vecDataFromFile[i]);
+        }
+        // todo
+
+        if (++i == numElements) break;
+    }
+    dataStruct.data = vecDataFromFile;
+    file.close();
+    return dataStruct;
+}
+
+template<typename T>
+DTensor<T> DTensor<T>::parseFromTextFile(std::string path_to_file,
+                                         StorageMode mode) {
+    auto parsedData = vectorFromFile<T>(path_to_file);
+    DTensor<T> tensorFromData(parsedData.data, parsedData.numRows, parsedData.numCols, parsedData.numMats);
+    return tensorFromData;
+}
+
+template<typename T>
+void DTensor<T>::saveToFile(std::string pathToFile) {
+    std::ofstream file(pathToFile);
+    file << numRows() << std::endl << numCols() << std::endl << numMats() << std::endl;
+    std::vector<T> myData(numEl()); download(myData);
+    if constexpr (std::is_floating_point<T>::value) {
+        int prec = std::numeric_limits<T>::max_digits10 - 1;
+        file << std::setprecision(prec);
+    }
+    for(const T& el : myData) file << el << std::endl;
+}
+
 template<typename T>
 void DTensor<T>::reshape(size_t newNumRows, size_t newNumCols, size_t newNumMats) {
     if (m_numRows == newNumRows && m_numCols == newNumCols && m_numMats == newNumMats) return;

main.cu

@@ -1,27 +1,14 @@
-#include <random>
-#include <algorithm>
-#include <iterator>
-#include <vector>
-#include <iostream>
-#include <cublas_v2.h>
 #include "include/tensor.cuh"
-#include <memory>
+#include <fstream>
+#include <iostream>
+#include <string>
+#include <vector>
 
-#define real_t double
 
 int main() {
-
-    std::vector<real_t> aData = {10.0, 2.0, 3.0,
-                            2.0, 20.0, -1.0,
-                            3.0, -1.0, 30.0};
-    DTensor<real_t> A(3, 3, 2);
-    DTensor<real_t> A0(A, 2, 0, 0);
-    DTensor<real_t> A1(A, 2, 1, 1);
-    A0.upload(aData);
-    A1.upload(aData);
-    CholeskyBatchFactoriser<real_t> chol(A);
-    chol.factorise();
-    std::cout << chol.info()(0);
-
+    auto z = DTensor<size_t>::parseFromTextFile("../test/data/my.dtensor",
+                                                StorageMode::rowMajor);
+    std::cout << z;
+    z.saveToFile("hohoho.dtensor");
     return 0;
 }

test/testTensor.cu

@@ -1,5 +1,6 @@
 #include <gtest/gtest.h>
 #include "../include/tensor.cuh"
+#include <filesystem>
 
 #define PRECISION_LOW 1e-4
 #define PRECISION_HIGH 1e-10
@@ -115,6 +116,30 @@ TEST_F(TensorTest, randomTensorCreation) {
     randomTensorCreation<int>();
 }
 
+/* ---------------------------------------
+ * Save to file and parse
+ * --------------------------------------- */
+
+TEMPLATE_WITH_TYPE_T
+void parseTensorFromFile() {
+    size_t nR = 20, nC = 40, nM = 60;
+    auto r = DTensor<T>::createRandomTensor(nR, nC, nM, -1, 1);
+    std::string fName = "myTest.dtensor";
+    r.saveToFile(fName);
+    auto a = DTensor<T>::parseFromTextFile(fName);
+    EXPECT_EQ(nR, a.numRows());
+    EXPECT_EQ(nC, a.numCols());
+    EXPECT_EQ(nM, a.numMats());
+    auto diff = a - r;
+    T err = diff.maxAbs();
+    EXPECT_LT(err, 2*std::numeric_limits<T>::epsilon());
+}
+
+TEST_F(TensorTest, parseTensorFromFile) {
+    parseTensorFromFile<float>();
+    parseTensorFromFile<double>();
+}
+
 /* ---------------------------------------
  * Move constructor
  * --------------------------------------- */