Implement device vector utils

src/kernels/matmul.cu

@@ -95,18 +95,6 @@ __global__ void Kernels::vec_exp(
 }
 
 
-__global__ void Kernels::clear(
-    float* __restrict__ d_vector,
-    const unsigned int w
-) {
-    int tid = blockDim.x * blockIdx.x + threadIdx.x;
-    if (tid >= w) {
-        return;
-    }
-    d_vector[tid] = 0.0f;
-}
-
-
 __global__ void Kernels::max_reduce(
     const float* __restrict__ d_vector,
     float* __restrict__ d_output,

src/layers/dense.cu

@@ -5,6 +5,7 @@
 #include <functional>
 #include <iostream>
 
+#include "vector.cuh"
 #include "activation.cuh"
 #include "cuda_helper.cuh"
 #include "dense.cuh"
@@ -63,6 +64,9 @@ void Dense::initializeBiases() {
 }
 
 float* Dense::forward(const float* d_input) {
+
+    CUDANet::Utils::clear(d_output, outputSize);
+
     Kernels::mat_vec_mul<<<forwardGridSize, BLOCK_SIZE>>>(
         d_weights, d_input, d_output, inputSize, outputSize
     );

src/utils/vector.cu

@@ -0,0 +1,58 @@
+#include <iostream>
+#include <vector>
+
+#include "vector.cuh"
+#include "matmul.cuh"
+#include "cuda_helper.cuh"
+
+using namespace CUDANet;
+
+void Utils::print_vec(float* d_vec, const unsigned int length) {
+    std::vector<float> h_vec(length);
+    CUDA_CHECK(cudaMemcpy(
+        h_vec.data(), d_vec, sizeof(float) * length, cudaMemcpyDeviceToHost
+    ));
+
+    for (int i = 0; i < length; ++i) {
+        std::cout << h_vec[i] << ", ";
+    }
+
+    std::cout << std::endl;
+}
+
+void Utils::clear(float* d_vec, const unsigned int length) {
+    CUDA_CHECK(cudaMemset(d_vec, 0, sizeof(float) * length));
+}
+
+void Utils::max(float* d_vec, float* d_max, const unsigned int length) {
+    
+    const int grid_size = (length + BLOCK_SIZE - 1) / BLOCK_SIZE;
+
+    Kernels::max_reduce<<<grid_size, BLOCK_SIZE>>>(d_vec, d_max, length);
+
+    int remaining = grid_size;
+    while (remaining > 1) {
+        int blocks_needed = (remaining + BLOCK_SIZE - 1) / BLOCK_SIZE;
+        CUDANet::Kernels::max_reduce<<<blocks_needed, BLOCK_SIZE>>>(d_max, d_max, remaining);
+        remaining = blocks_needed;
+    }
+
+}
+
+void Utils::sum(float* d_vec, float* d_sum, const unsigned int length) {
+    
+    const int gridSize = (length + BLOCK_SIZE - 1) / BLOCK_SIZE;
+
+
+    CUDANet::Kernels::sum_reduce<<<gridSize, BLOCK_SIZE>>>(
+        d_vec, d_sum, length
+    );
+
+    int remaining = gridSize;
+    while (remaining > 1) {
+        std::cout << remaining << std::endl;
+        int blocks_needed = (remaining + BLOCK_SIZE - 1) / BLOCK_SIZE;
+        CUDANet::Kernels::sum_reduce<<<blocks_needed, BLOCK_SIZE>>>(d_sum, d_sum, remaining);
+        remaining = blocks_needed;
+    }
+}