Upload New File

cudaLenet5.cu

+#include <chrono>
+#include <iostream>
+#include <cuda_runtime.h>
+#include <cuda_profiler_api.h>
+#include <device_launch_parameters.h>
+#include <assert.h>
+#include <stdio.h>
+#include "tensor.h"
+
+/*__global__ void conv1Kernel(){
+    int i = threadIdx.x;
+    printf("Thread i: %d\n", i);
+    __syncthreads();
+}
+*/
+// Input 28x28x6xN -> Output 14x14x6xN;
+__global__ void pool1Kernel(float* data, float* res){
+    int i = threadIdx.x;
+    printf("Thread i: %d\n", i);
+
+    __syncthreads();
+}
+
+/*
+__global__ void conv2Kernel(){
+    int i = threadIdx.x;
+    printf("Thread i: %d\n", i);
+    __syncthreads();
+}
+
+__global__ void pool2Kernel(){
+    int i = threadIdx.x;
+    printf("Thread i: %d\n", i);
+    __syncthreads();
+}
+
+__global__ void conv3Kernel(){
+    int i = threadIdx.x;
+    printf("Thread i: %d\n", i);
+    __syncthreads();
+}
+
+__global__ void fully1Kernel(){
+    int i = threadIdx.x;
+    printf("Thread i: %d\n", i);
+    __syncthreads();
+}
+
+__global__ void fully2Kernel(){
+    int i = threadIdx.x;
+    printf("Thread i: %d\n", i);
+    __syncthreads();
+}
+*/
+//Tensor Bilder 4D (32x32xM), ... 
+cudaError_t lenet5Cuda(){
+    // Setup (vorerst als float pointer)
+    //int N = 128;
+    cudaError_t cudaStatus;
+    cudaStatus = cudaSetDevice(0);
+    if (cudaStatus != cudaSuccess) {
+        fprintf(stderr, "lenet5Cuda: cudaSetDevice failed!  Do you have a CUDA-capable GPU installed?\n");
+        return cudaStatus;
+    }
+    
+    /*
+    // Memory for processed Data after each step
+    float *dev_images, *dev_conv1, *dev_pool1, *dev_conv2, *dev_pool2, *dev_conv3, *dev_fully1, *dev_fully2;
+    // Memory for wheigts which were used to process
+    // Anzahl:  156,        2416,       48120,        10164,        850;
+    /*float *dev_w_conv1, *dev_w_conv2, *dev_w_conv3, *dev_w_fully1, *dev_w_fully2;
+    cudaStatus = cudaMalloc((void**)&dev_images, 32*32*M*sizeof(float));
+    cudaStatus = cudaMalloc((void**)&dev_conv1, 28*28*6*N*sizeof(float));
+    cudaStatus = cudaMalloc((void**)&dev_pool1, 14*14*6*N*sizeof(float));
+    cudaStatus = cudaMalloc((void**)&dev_conv2, 10*10*16*N*sizeof(float));
+    cudaStatus = cudaMalloc((void**)&dev_pool2, 5*5*16*N*sizeof(float));
+    cudaStatus = cudaMalloc((void**)&dev_conv3, 120*N*sizeof(float));
+    cudaStatus = cudaMalloc((void**)&dev_fully1, 84*N*sizeof(float));
+    cudaStatus = cudaMalloc((void**)&dev_fully2, 10*M*sizeof(float));
+
+    cudaStatus = cudaMalloc((void**)&dev_w_conv1, (5*5*6+6)*sizeof(float));
+    cudaStatus = cudaMalloc((void**)&dev_w_conv2, (5*5*6*16+16)*sizeof(float));
+    cudaStatus = cudaMalloc((void**)&dev_w_conv3, (5*5*16*120+120)*sizeof(float));
+    cudaStatus = cudaMalloc((void**)&dev_w_fully1, (120*84+84)*sizeof(float));
+    cudaStatus = cudaMalloc((void**)&dev_w_fully2, (84*10+10)sizeof(float));
+
+    // Input 32x32xN; N = #Bilder
+    // ->Conv2D(tanh)->
+
+    // Output 28x28x6xN; N = #Bilder
+
+    // Input 28x28x6xN; N = #Bilder
+    // ->Avg-Pooling(2x2)-> 
+    pool1Kernel<<<1, 10>>>(dev_conv1, dev_pool1);
+    if (cudaStatus != cudaSuccess) {
+        fprintf(stderr, "pool1Cuda: pool1 launch failed: %s\n", cudaGetErrorString(cudaStatus));  
+        return cudaStatus;
+    }
+    // Output 14x14x6xN; N = #Bilder
+
+    // Input 14x14x6xN; N = #Bilder
+    // ->Conv2D(tanh)->
+
+    // Output 10x10x16xN; N = #Bilder
+
+    // Input 10x10x16xN; N = #Bilder
+    // ->Avg-Pooling(2x2)->
+
+    // Output 5x5x16xN; N = #Bilder
+
+    // Input 5x5x16xN; N = #Bilder
+    // ->Conv2D(tanh)->
+
+    // Output 120xN; N = #Bilder
+
+    // Input 120xN; N = #Bilder
+    // ->FullyConnected->
+
+    // Output 84xN; N = #Bilder
+
+    // Input 84xN; N = #Bilder
+    // ->FullyConnected->
+
+    // Output 10xN; N = #Bilder
+    */
+    return cudaStatus;
+}