Remove cublas dependency

src/layers/dense.cu

@@ -1,4 +1,3 @@
-#include <cublas_v2.h>
 #include <cuda_runtime.h>
 
 #include <cstdio>
@@ -9,16 +8,15 @@
 #include "activations.cuh"
 #include "cuda_helper.cuh"
 #include "dense.cuh"
+#include "matrix_math.cuh"
 
 Layers::Dense::Dense(
     int            inputSize,
     int            outputSize,
-    Activation     activation,
-    cublasHandle_t cublasHandle
+    Activation     activation
 )
     : inputSize(inputSize),
       outputSize(outputSize),
-      cublasHandle(cublasHandle),
       activation(activation) {
     // Allocate memory for weights and biases
     weights.resize(outputSize * inputSize);
@@ -54,35 +52,30 @@ void Layers::Dense::initializeBiases() {
 }
 
 void Layers::Dense::forward(const float* d_input, float* d_output) {
-    const float alpha = 1.0f;
-    const float beta  = 0.0f;
 
-    CUBLAS_CHECK(cublasSgemv(
-        cublasHandle, CUBLAS_OP_N, outputSize, inputSize, &alpha, d_weights,
-        outputSize, d_input, 1, &beta, d_output, 1
-    ));
-    CUBLAS_CHECK(
-        cublasSaxpy(cublasHandle, outputSize, &alpha, d_biases, 1, d_output, 1)
+    mat_vec_mul_kernel<<<1, outputSize>>>(
+        d_weights, d_input, d_output, inputSize, outputSize
     );
 
-    int threadsPerBlock = 256;
-    int blocksPerGrid   = (outputSize + threadsPerBlock - 1) / threadsPerBlock;
+    vec_vec_add_kernel<<<1, outputSize>>>(
+        d_biases, d_output, d_output, outputSize
+    );
 
     switch (activation) {
         case SIGMOID:
-            sigmoid_kernel<<<blocksPerGrid, threadsPerBlock>>>(
+            sigmoid_kernel<<<1, outputSize>>>(
                 d_output, d_output, outputSize
             );
             break;
 
         case RELU:
-            relu_kernel<<<blocksPerGrid, threadsPerBlock>>>(
+            relu_kernel<<<1, outputSize>>>(
                 d_output, d_output, outputSize
             );
             break;
 
         default:
-            linear_kernel<<<blocksPerGrid, threadsPerBlock>>>(
+            linear_kernel<<<1, outputSize>>>(
                 d_output, d_output, outputSize
             );
             break;
@@ -92,12 +85,13 @@ void Layers::Dense::forward(const float* d_input, float* d_output) {
 }
 
 void Layers::Dense::toCuda() {
-    CUBLAS_CHECK(cublasSetMatrix(
-        outputSize, inputSize, sizeof(float), weights.data(), outputSize,
-        d_weights, outputSize
+    CUDA_CHECK(cudaMemcpy(
+        d_weights, weights.data(), sizeof(float) * inputSize * outputSize,
+        cudaMemcpyHostToDevice
     ));
-    CUBLAS_CHECK(cublasSetVector(
-        biases.size(), sizeof(float), biases.data(), 1, d_biases, 1
+    CUDA_CHECK(cudaMemcpy(
+        d_biases, biases.data(), sizeof(float) * outputSize,
+        cudaMemcpyHostToDevice
     ));
 }
 
@@ -111,10 +105,9 @@ void Layers::Dense::setWeights(
         exit(EXIT_FAILURE);
     }
 
-    for (int j = 0; j < inputSize; ++j) {
-        for (int i = 0; i < outputSize; ++i) {
-            int idx      = IDX2C(i, j, outputSize);
-            weights[idx] = weights_input[i][j];
+    for (int i = 0; i < outputSize; ++i) {
+        for (int j = 0; j < inputSize; ++j) {
+            weights[i * inputSize + j] = weights_input[i][j];
         }
     }