Hi all,

I have a simple cuda program which executed on the CPU and on the GPU, gives different results when the GPU is a Tesla card and gives the same results when the GPU is a Fermi card. 

Does anybody know any explanation for the difference of the results?

I post also the code of my program:

/* File: testDefinitions.h */ #define MVMul(v1, m, v2) \ (v1).x = (m)[0] * (v2).x + (m)[3] * (v2).y + (m)[6] * (v2).z, \ (v1).y = (m)[1] * (v2).x + (m)[4] * (v2).y + (m)[7] * (v2).z, \ (v1).z = (m)[2] * (v2).x + (m)[5] * (v2).y + (m)[8] * (v2).z typedef struct { float x, y, z; } triple; typedef struct { float u[9]; } ninuple; /* File testc.c */ #include #include #include #include "testDefinitions.h" float uNewGpuHost[3]; void preparation_gpu_matrix_product(triple, ninuple); // main routine that executes on the host int main(int argc, char **argv) { triple uOld, uNewGpu, uNewHost; ninuple Ru; uOld.x = 2.39282537; uOld.y = 0.653135121; uOld.z = -0.224300578; Ru.u[0] = 1; Ru.u[1] = 0; Ru.u[2] = 0; Ru.u[3] = 0; Ru.u[4] = 0.999999523; Ru.u[5] = -0.000974848808; Ru.u[6] = 0; Ru.u[7] = 0.000974848808; Ru.u[8] = 0.999999523; MVMul(uNewHost, Ru.u, uOld); printf("\nValue of uNew calculated by host: %.20f %.20f %.20f\n", uNewHost.x, uNewHost.y, uNewHost.z); preparation_gpu_matrix_product(uOld, Ru); uNewGpu.x = uNewGpuHost[0]; uNewGpu.y = uNewGpuHost[1]; uNewGpu.z = uNewGpuHost[2]; printf("Value of uNew calculated by GPU: %.20f %.20f %.20f\n", uNewGpu.x, uNewGpu.y, uNewGpu.z); if(((uNewHost.x-uNewGpu.x)==0) && ((uNewHost.y-uNewGpu.y)==0) && ((uNewHost.z-uNewGpu.z)==0)) { printf("Test PASSED\n"); } else printf("Test FAILED\n"); return 0; } /* File test.cu */ #include #include __global__ void matrix_product(triple, ninuple, float*); extern float uNewGpuHost[3]; extern "C" { void preparation_gpu_matrix_product(triple uOld, ninuple Ru) { float* uNewDev; cudaMalloc((void **) &uNewDev, 3*sizeof(float)); // Allocate array on device cudaError_t error=cudaGetLastError(); if(error != cudaSuccess)printf("cudaMalloc: %s\n",cudaGetErrorString(error)); // Do calculation on device: matrix_product <<< 1, 16 >>> (uOld, Ru, uNewDev); cudaThreadSynchronize(); error=cudaGetLastError(); if(error != cudaSuccess)printf("Synchronize kernel: %s\n",cudaGetErrorString(error)); // Retrieve result from device and store it in host array cudaMemcpy(uNewGpuHost, uNewDev, sizeof(float)*3, cudaMemcpyDeviceToHost); error=cudaGetLastError(); if(error != cudaSuccess)printf("cuda copy device to host: %s\n",cudaGetErrorString(error)); cudaFree(uNewDev); } } __global__ void matrix_product(triple uOld, ninuple Ru, float* uNewDev) { int idx = blockIdx.x * blockDim.x + threadIdx.x; if(idx==0) { triple uNew; MVMul(uNew, Ru.u, uOld); uNewDev[0]=uNew.x; uNewDev[1]=uNew.y; uNewDev[2]=uNew.z; } } The results of this simple program targeted to a Tesla card: x295 are: Value of uNew calculated by host: 2.39282536506652832031 0.65291619300842285156 -0.22493718564510345459 Value of uNew calculated by GPU: 2.39282536506652832031 0.65291619300842285156 -0.22493717074394226074 Test FAILED insted when run on a Fermi card the results are: Value of uNew calculated by host: 2.39282536506652832031 0.65291619300842285156 -0.22493718564510345459 Value of uNew calculated by GPU: 2.39282536506652832031 0.65291619300842285156 -0.22493718564510345459 Test PASSED Does anybody know why? Thank you in advance, Ardita