getting rid of sub_arrays, nvidia driver memory leak

Joseph Redmon

2014-10-28 af4e4f92dc9e5da160eb6c6870a7b38b863f1c6c

 src/convolutional_layer.c

@@ -369,11 +369,9 @@

    for(i = 0; i < layer.batch; ++i){
        cl_mem a = layer.filters_cl;
        cl_mem b = cl_sub_array(layer.col_image_cl, i*k*n, k*n);
        cl_mem c = cl_sub_array(layer.output_cl, i*m*n, m*n);
        gemm_ongpu(0,0,m,n,k,1.,a,k,b,n,1.,c,n);
        clReleaseMemObject(b);
        clReleaseMemObject(c);
        cl_mem b = layer.col_image_cl;
        cl_mem c = layer.output_cl;
        gemm_ongpu_offset(0,0,m,n,k,1.,a,0,k,b,i*k*n,n,1.,c,i*m*n,n);
    }
    #ifdef TIMEIT
    clFinish(cl.queue);
@@ -396,14 +394,11 @@
    learn_bias_convolutional_layer_ongpu(layer);

    for(i = 0; i < layer.batch; ++i){
        cl_mem a = cl_sub_array(layer.delta_cl,i*m*k, m*k);
        cl_mem b = cl_sub_array(layer.col_image_cl,i*k*n, k*n);
        cl_mem a = layer.delta_cl;
        cl_mem b = layer.col_image_cl;
        cl_mem c = layer.filter_updates_cl;

        gemm_ongpu(0,1,m,n,k,1,a,k,b,k,1,c,n);

        clReleaseMemObject(a);
        clReleaseMemObject(b);
        gemm_ongpu_offset(0,1,m,n,k,1,a,i*m*k,k,b,i*k*n,k,1,c,0,n);
    }
    //cl_read_array(layer.delta_cl, layer.delta, m*k*layer.batch);

@@ -415,12 +410,10 @@

        for(i = 0; i < layer.batch; ++i){
            cl_mem a = layer.filters_cl;
            cl_mem b = cl_sub_array(layer.delta_cl, i*k*n, k*n);
            cl_mem c = cl_sub_array(layer.col_image_cl, i*m*n, m*n);
            cl_mem b = layer.delta_cl;
            cl_mem c = layer.col_image_cl;

            gemm_ongpu(1,0,m,n,k,1,a,m,b,n,0,c,n);
            clReleaseMemObject(b);
            clReleaseMemObject(c);
            gemm_ongpu_offset(1,0,m,n,k,1,a,0,m,b,i*k*n,n,0,c,i*m*n,n);
        }

        scal_ongpu(layer.batch*layer.h*layer.w*layer.c,0,delta_cl, 1);