Improve cuda error handling

src/layers/activation.cu

@@ -38,10 +38,12 @@ void Activation::activate(float* d_input) {
             Kernels::sigmoid<<<gridSize, BLOCK_SIZE>>>(
                 d_input, d_input, length
             );
+            CUDA_CHECK(cudaGetLastError());
             break;
 
         case RELU:
             Kernels::relu<<<gridSize, BLOCK_SIZE>>>(d_input, d_input, length);
+            CUDA_CHECK(cudaGetLastError());
             break;
         case SOFTMAX:
 
@@ -52,11 +54,13 @@ void Activation::activate(float* d_input) {
             Kernels::vec_scalar_sub<<<gridSize, BLOCK_SIZE>>>(
                 d_input, d_input, d_max, length
             );
+            CUDA_CHECK(cudaGetLastError());
 
             // Compute exponentials
             Kernels::vec_exp<<<gridSize, BLOCK_SIZE>>>(
                 d_input, d_input, length
             );
+            CUDA_CHECK(cudaGetLastError());
 
             // Find sum
             Utils::sum(d_input, d_softmax_sum, length);
@@ -64,6 +68,7 @@ void Activation::activate(float* d_input) {
             Kernels::vec_scalar_div<<<gridSize, BLOCK_SIZE>>>(
                 d_input, d_input, d_softmax_sum, length
             );
+            CUDA_CHECK(cudaGetLastError());
 
             break;
 
@@ -71,6 +76,6 @@ void Activation::activate(float* d_input) {
             break;    
     }
 
-    cudaDeviceSynchronize();
+    CUDA_CHECK(cudaDeviceSynchronize());
 }
 

src/layers/add.cu

@@ -25,5 +25,7 @@ void Add::forward(const float* d_inputA, const float* d_inputB) {
     Kernels::vec_vec_add<<<gridSize, BLOCK_SIZE>>>(
         d_inputA, d_inputB, d_output, inputSize
     );
+    CUDA_CHECK(cudaGetLastError());
+    CUDA_CHECK(cudaDeviceSynchronize());
 
 }

src/layers/avg_pooling.cu

@@ -42,8 +42,10 @@ float* AvgPooling2D::forward(const float* d_input) {
     Kernels::avg_pooling<<<grid, block>>>(
         d_input, d_output, inputSize, outputSize, nChannels, poolingSize, stride
     );
+    CUDA_CHECK(cudaGetLastError());
 
     activation.activate(d_output);
+    CUDA_CHECK(cudaDeviceSynchronize());
 
     return d_output;
 }

src/layers/concat.cu

@@ -23,10 +23,13 @@ float* Concat::forward(const float* d_input_A, const float* d_input_B) {
     CUDA_CHECK(cudaMemcpy(
         d_output, d_input_A, sizeof(float) * inputASize, cudaMemcpyDeviceToDevice
     ));
+
     CUDA_CHECK(cudaMemcpy(
         d_output + inputASize, d_input_B,
         sizeof(float) * inputBSize, cudaMemcpyDeviceToDevice
     ));
 
+    CUDA_CHECK(cudaDeviceSynchronize());
+
     return d_output;
 }

src/layers/conv2d.cu

@@ -3,6 +3,11 @@
 #include "convolution.cuh"
 #include "cuda_helper.cuh"
 #include "matmul.cuh"
+#include "layer.cuh"
+#include "vector.cuh"
+
+#include <iostream>
+#include <vector>
 
 using namespace CUDANet::Layers;
 
@@ -100,6 +105,7 @@ void Conv2d::toCuda() {
 }
 
 float* Conv2d::forward(const float* d_input) {
+
     // Convolve
     dim3 block(8,8,8);
     dim3 grid(
@@ -108,11 +114,14 @@ float* Conv2d::forward(const float* d_input) {
         (numFilters + block.z - 1) / block.z
     );
 
+    CUDANet::Utils::clear(d_output, outputSize * outputSize * numFilters);
+
     Kernels::convolution<<<grid, block>>>(
         d_input, d_weights, d_biases, d_output, inputSize, inputChannels, paddingSize,
         kernelSize, stride, numFilters, outputSize
     );
-
+    CUDA_CHECK(cudaGetLastError());
+ 
     // Apply activation
     activation.activate(d_output);
 

src/layers/dense.cu

@@ -65,18 +65,23 @@ void Dense::initializeBiases() {
 
 float* Dense::forward(const float* d_input) {
 
-    CUDANet::Utils::clear(d_output, outputSize);
+    // CUDANet::Utils::clear(d_output, outputSize);
+
+    // CUDA_CHECK(cudaPeekAtLastError());
+
+    std::cout << "Dense::forward" << std::endl;
 
     Kernels::mat_vec_mul<<<forwardGridSize, BLOCK_SIZE>>>(
         d_weights, d_input, d_output, inputSize, outputSize
     );
+    CUDA_CHECK(cudaPeekAtLastError());
 
     Kernels::vec_vec_add<<<biasGridSize, BLOCK_SIZE>>>(
         d_biases, d_output, d_output, outputSize
     );
+    CUDA_CHECK(cudaPeekAtLastError());
 
     activation.activate(d_output);
-
     CUDA_CHECK(cudaDeviceSynchronize());
 
     return d_output;

src/layers/input.cu

@@ -12,17 +12,11 @@ Input::~Input() {
     cudaFree(d_output);
 }
 
-/*
-Copies host input to device d_output
-
-Args
-    const float* input Host pointer to input data
-    float* d_output Device pointer to input data copied to device
-*/
 float* Input::forward(const float* input) {
     CUDA_CHECK(cudaMemcpy(
         d_output, input, sizeof(float) * inputSize, cudaMemcpyHostToDevice
     ));
+    CUDA_CHECK(cudaDeviceSynchronize());
 
     return d_output;
 }

src/layers/max_pooling.cu

@@ -45,8 +45,10 @@ float* MaxPooling2D::forward(const float* d_input) {
     Kernels::max_pooling<<<grid, block>>>(
         d_input, d_output, inputSize, outputSize, nChannels, poolingSize, stride
     );
+    CUDA_CHECK(cudaGetLastError());
 
     activation.activate(d_output);
+    CUDA_CHECK(cudaDeviceSynchronize());
 
     return d_output;
 }

src/layers/output.cu

@@ -17,6 +17,7 @@ float* Output::forward(const float* input) {
     CUDA_CHECK(cudaMemcpy(
         h_output, input, sizeof(float) * inputSize, cudaMemcpyDeviceToHost
     ));
+    CUDA_CHECK(cudaDeviceSynchronize());
 
     return h_output;
 }

src/utils/vector.cu

@@ -29,11 +29,14 @@ 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);
+    CUDA_CHECK(cudaGetLastError());
 
     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);
+        CUDA_CHECK(cudaGetLastError());
+
         remaining = blocks_needed;
     }
 
@@ -47,11 +50,14 @@ void Utils::sum(float* d_vec, float* d_sum, const unsigned int length) {
     CUDANet::Kernels::sum_reduce<<<gridSize, BLOCK_SIZE>>>(
         d_vec, d_sum, length
     );
+    CUDA_CHECK(cudaGetLastError());
 
     int remaining = gridSize;
     while (remaining > 1) {
         int blocks_needed = (remaining + BLOCK_SIZE - 1) / BLOCK_SIZE;
         CUDANet::Kernels::sum_reduce<<<blocks_needed, BLOCK_SIZE>>>(d_sum, d_sum, remaining);
+        CUDA_CHECK(cudaGetLastError());
+        
         remaining = blocks_needed;
     }
 }