Initialize and use cuBLAS properly

include/layers/dense.h

@@ -15,7 +15,8 @@ namespace Layers {
         ~Dense();
 
         void forward(const float* input, float* output);
-        void to_cuda();
+        virtual void setWeights(const std::vector<std::vector<float>>& weights) = 0;
+        virtual void setBiases(const std::vector<float>& biases) = 0;
 
     private:
         int inputSize;
@@ -31,6 +32,7 @@ namespace Layers {
 
         void initializeWeights();
         void initializeBiases();
+        void toCuda();
     };
 
 } // namespace Layers

include/layers/ilayer.h

@@ -11,7 +11,6 @@ namespace Layers {
         virtual ~ILayer() {}
 
         virtual void forward(const float* input, float* output) = 0;
-        virtual void to_cuda() = 0;
     };
 
 } // namespace Layers

include/utils/cuda_helper.h

@@ -2,6 +2,7 @@
 #define CUDA_HELPER_H
 
 #include <cuda_runtime.h>
+#include <cublas_v2.h>
 
 #define IDX2C(i,j,ld) (((j)*(ld))+(i))
 
@@ -17,7 +18,15 @@ do { \
     } \
 } while (0)
 
-// Initialize CUDA and return the device properties
+// cuBLAS error checking macro
-cudaDeviceProp initializeCUDA();
+#define CUBLAS_CHECK(call) \
+do { \
+    cublasStatus_t result = call; \
+    if (result != CUBLAS_STATUS_SUCCESS) { \
+        fprintf(stderr, "cuBLAS error at %s:%d code=%d\n", \
+                __FILE__, __LINE__, static_cast<int>(result)); \
+        exit(EXIT_FAILURE); \
+    } \
+} while (0)
 
 #endif // CUDA_HELPER_H

src/layers/dense.cpp

@@ -21,7 +21,7 @@ Layers::Dense::Dense(int inputSize, int outputSize, cublasHandle_t cublasHandle)
     CUDA_CHECK(cudaMalloc((void**)&d_weights, sizeof(float) * inputSize * outputSize));
     CUDA_CHECK(cudaMalloc((void**)&d_biases, sizeof(float) * biases.size()));
 
-    to_cuda();
+    toCuda();
 }
 
 Layers::Dense::~Dense() {
@@ -54,7 +54,17 @@ void Layers::Dense::forward(const float* input, float* output) {
     cublasSaxpy(cublasHandle, outputSize, &alpha, d_biases, 1, output, 1);
 }
 
-void Layers::Dense::to_cuda() {
+void Layers::Dense::toCuda() {
-    CUDA_CHECK(cudaMemcpy(d_weights, weights.data(), sizeof(float) * inputSize * outputSize, cudaMemcpyHostToDevice));
+    CUBLAS_CHECK(cublasSetMatrix(outputSize, inputSize, sizeof(float), weights.data(), inputSize, d_weights, outputSize));
-    CUDA_CHECK(cudaMemcpy(d_biases, biases.data(), sizeof(float) * biases.size(), cudaMemcpyHostToDevice));
+    CUBLAS_CHECK(cublasSetVector(biases.size(), sizeof(float), biases.data(), 1, d_biases, 1));
+}
+
+void Layers::Dense::setWeights(const std::vector<std::vector<float>>& weights) {
+    this->weights = weights;
+    toCuda();
+}
+
+void Layers::Dense::setBiases(const std::vector<float>& biases) {
+    this->biases = biases;
+    toCuda();
 }

src/utils/cuda_helper.cpp

@@ -1,9 +1,9 @@
 #include <cstdio>
 #include <cstdlib>
 #include "cuda_helper.h"
+#include <cublas_v2.h>
 
-// Initialize CUDA and return the device properties
+cudaDeviceProp initializeCUDA(cublasHandle_t& cublasHandle) {
-cudaDeviceProp initializeCUDA() {
     int deviceCount;
     CUDA_CHECK(cudaGetDeviceCount(&deviceCount));
 
@@ -12,7 +12,7 @@ cudaDeviceProp initializeCUDA() {
         std::exit(EXIT_FAILURE);
     }
 
-    int device = 0; // You can modify this to choose a different GPU
+    int device = 0;
     CUDA_CHECK(cudaSetDevice(device));
 
     cudaDeviceProp deviceProp;
@@ -20,5 +20,8 @@ cudaDeviceProp initializeCUDA() {
 
     std::printf("Using CUDA device %d: %s\n", device, deviceProp.name);
 
+    // Initialize cuBLAS
+    CUBLAS_CHECK(cublasCreate(&cublasHandle));
+
     return deviceProp;
-}
+}