static size_t get_workspace_size(layer l){
#ifdef CUDNN
if(gpu_index >= 0){
size_t most = 0;
size_t s = 0;
cudnnGetConvolutionForwardWorkspaceSize(cudnn_handle(),
l.srcTensorDesc,
l.weightDesc,
l.convDesc,
l.dstTensorDesc,
l.fw_algo,
&s);
if (s > most) most = s;
cudnnGetConvolutionBackwardFilterWorkspaceSize(cudnn_handle(),
l.srcTensorDesc,
l.ddstTensorDesc,
l.convDesc,
l.dweightDesc,
l.bf_algo,
&s);
if (s > most) most = s;
cudnnGetConvolutionBackwardDataWorkspaceSize(cudnn_handle(),
l.weightDesc,
l.ddstTensorDesc,
l.convDesc,
l.dsrcTensorDesc,
l.bd_algo,
&s);
if (s > most) most = s;
return most;
}
#endif
return (size_t)l.out_h*l.out_w*l.size*l.size*l.c/l.groups*sizeof(float);
}
size_t get_workspace_size(layer l){
#ifdef CUDNN
size_t most = 0;
size_t s = 0;
cudnnGetConvolutionForwardWorkspaceSize(cudnn_handle(),
l.srcTensorDesc,
l.filterDesc,
l.convDesc,
l.dstTensorDesc,
l.fw_algo,
&s);
if (s > most) most = s;
cudnnGetConvolutionBackwardFilterWorkspaceSize(cudnn_handle(),
l.srcTensorDesc,
l.ddstTensorDesc,
l.convDesc,
l.dfilterDesc,
l.bf_algo,
&s);
if (s > most) most = s;
cudnnGetConvolutionBackwardDataWorkspaceSize(cudnn_handle(),
l.filterDesc,
l.ddstTensorDesc,
l.convDesc,
l.dsrcTensorDesc,
l.bd_algo,
&s);
if (s > most) most = s;
return most;
#else
return (size_t)l.out_h*l.out_w*l.size*l.size*l.c*sizeof(float);
#endif
}
void cudnn_convolutional_setup(layer *l)
{
cudnnSetTensor4dDescriptor(l->dsrcTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, l->batch, l->c, l->h, l->w);
cudnnSetTensor4dDescriptor(l->ddstTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, l->batch, l->out_c, l->out_h, l->out_w);
cudnnSetTensor4dDescriptor(l->srcTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, l->batch, l->c, l->h, l->w);
cudnnSetTensor4dDescriptor(l->dstTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, l->batch, l->out_c, l->out_h, l->out_w);
cudnnSetTensor4dDescriptor(l->normTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, 1, l->out_c, 1, 1);
cudnnSetFilter4dDescriptor(l->dweightDesc, CUDNN_DATA_FLOAT, CUDNN_TENSOR_NCHW, l->n, l->c/l->groups, l->size, l->size);
cudnnSetFilter4dDescriptor(l->weightDesc, CUDNN_DATA_FLOAT, CUDNN_TENSOR_NCHW, l->n, l->c/l->groups, l->size, l->size);
#if CUDNN_MAJOR >= 6
cudnnSetConvolution2dDescriptor(l->convDesc, l->pad, l->pad, l->stride, l->stride, 1, 1, CUDNN_CROSS_CORRELATION, CUDNN_DATA_FLOAT);
#else
cudnnSetConvolution2dDescriptor(l->convDesc, l->pad, l->pad, l->stride, l->stride, 1, 1, CUDNN_CROSS_CORRELATION);
#endif
#if CUDNN_MAJOR >= 7
cudnnSetConvolutionGroupCount(l->convDesc, l->groups);
#else
if(l->groups > 1){
error("CUDNN < 7 doesn't support groups, please upgrade!");
}
#endif
cudnnGetConvolutionForwardAlgorithm(cudnn_handle(),
l->srcTensorDesc,
l->weightDesc,
l->convDesc,
l->dstTensorDesc,
CUDNN_CONVOLUTION_FWD_PREFER_FASTEST,
0,
&l->fw_algo);
cudnnGetConvolutionBackwardDataAlgorithm(cudnn_handle(),
l->weightDesc,
l->ddstTensorDesc,
l->convDesc,
l->dsrcTensorDesc,
CUDNN_CONVOLUTION_BWD_DATA_PREFER_FASTEST,
0,
&l->bd_algo);
cudnnGetConvolutionBackwardFilterAlgorithm(cudnn_handle(),
l->srcTensorDesc,
l->ddstTensorDesc,
l->convDesc,
l->dweightDesc,
CUDNN_CONVOLUTION_BWD_FILTER_PREFER_FASTEST,
0,
&l->bf_algo);
}
void Activation::compute_gpu(const vector<bool>& add) {
DTYPE alpha = 1.;
DTYPE beta = add[0] ? 1. : 0.;
CUDNN_CHECK(cudnnActivationForward(cudnn_handle(), activation_mode_,
&alpha, bottom_desc_, inputs_[0]->gpu_data(), &beta, top_desc_,
outputs_[0]->mutable_gpu_data()));
}
void Softmax::compute_gpu(const vector<bool>& add) {
DTYPE alpha = 1.;
DTYPE beta = add[0] ? 1. : 0.;
CUDNN_CHECK(cudnnSoftmaxForward(cudnn_handle(), CUDNN_SOFTMAX_ACCURATE,
softmax_mode_, &alpha, bottom_desc_, inputs_[0]->gpu_data(), &beta,
top_desc_, outputs_[0]->mutable_gpu_data()));
}
void forward_batchnorm_layer_gpu(layer l, network_state state)
{
if(l.type == BATCHNORM) copy_ongpu(l.outputs*l.batch, state.input, 1, l.output_gpu, 1);
if(l.type == CONNECTED){
l.out_c = l.outputs;
l.out_h = l.out_w = 1;
}
if (state.train) {
#ifdef CUDNN
copy_ongpu(l.outputs*l.batch, l.output_gpu, 1, l.x_gpu, 1);
float one = 1;
float zero = 0;
cudnnBatchNormalizationForwardTraining(cudnn_handle(),
CUDNN_BATCHNORM_SPATIAL,
&one,
&zero,
l.dstTensorDesc,
l.x_gpu,
l.dstTensorDesc,
l.output_gpu,
l.normTensorDesc,
l.scales_gpu,
l.biases_gpu,
.01,
l.rolling_mean_gpu,
l.rolling_variance_gpu,
.00001,
l.mean_gpu,
l.variance_gpu);
#else
fast_mean_gpu(l.output_gpu, l.batch, l.out_c, l.out_h*l.out_w, l.mean_gpu);
fast_variance_gpu(l.output_gpu, l.mean_gpu, l.batch, l.out_c, l.out_h*l.out_w, l.variance_gpu);
scal_ongpu(l.out_c, .99, l.rolling_mean_gpu, 1);
axpy_ongpu(l.out_c, .01, l.mean_gpu, 1, l.rolling_mean_gpu, 1);
scal_ongpu(l.out_c, .99, l.rolling_variance_gpu, 1);
axpy_ongpu(l.out_c, .01, l.variance_gpu, 1, l.rolling_variance_gpu, 1);
copy_ongpu(l.outputs*l.batch, l.output_gpu, 1, l.x_gpu, 1);
normalize_gpu(l.output_gpu, l.mean_gpu, l.variance_gpu, l.batch, l.out_c, l.out_h*l.out_w);
copy_ongpu(l.outputs*l.batch, l.output_gpu, 1, l.x_norm_gpu, 1);
scale_bias_gpu(l.output_gpu, l.scales_gpu, l.batch, l.out_c, l.out_h*l.out_w);
add_bias_gpu(l.output_gpu, l.biases_gpu, l.batch, l.out_c, l.out_w*l.out_h);
#endif
} else {
normalize_gpu(l.output_gpu, l.rolling_mean_gpu, l.rolling_variance_gpu, l.batch, l.out_c, l.out_h*l.out_w);
scale_bias_gpu(l.output_gpu, l.scales_gpu, l.batch, l.out_c, l.out_h*l.out_w);
add_bias_gpu(l.output_gpu, l.biases_gpu, l.batch, l.out_c, l.out_w*l.out_h);
}
}
void cudnn_convolutional_setup(layer *l)
{
cudnnSetTensor4dDescriptor(l->dsrcTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, l->batch, l->c, l->h, l->w);
cudnnSetTensor4dDescriptor(l->ddstTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, l->batch, l->out_c, l->out_h, l->out_w);
cudnnSetFilter4dDescriptor(l->dweightDesc, CUDNN_DATA_FLOAT, CUDNN_TENSOR_NCHW, l->n, l->c, l->size, l->size);
cudnnSetTensor4dDescriptor(l->srcTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, l->batch, l->c, l->h, l->w);
cudnnSetTensor4dDescriptor(l->dstTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, l->batch, l->out_c, l->out_h, l->out_w);
cudnnSetFilter4dDescriptor(l->weightDesc, CUDNN_DATA_FLOAT, CUDNN_TENSOR_NCHW, l->n, l->c, l->size, l->size);
//cudnnSetConvolution2dDescriptor(l->convDesc, l->pad, l->pad, l->stride, l->stride, 1, 1, CUDNN_CROSS_CORRELATION);
cudnnSetConvolution2dDescriptor(l->convDesc, l->pad, l->pad, l->stride, l->stride, 1, 1, CUDNN_CROSS_CORRELATION, CUDNN_DATA_FLOAT);
cudnnGetConvolutionForwardAlgorithm(cudnn_handle(),
l->srcTensorDesc,
l->weightDesc,
l->convDesc,
l->dstTensorDesc,
CUDNN_CONVOLUTION_FWD_PREFER_FASTEST,
0,
&l->fw_algo);
cudnnGetConvolutionBackwardDataAlgorithm(cudnn_handle(),
l->weightDesc,
l->ddstTensorDesc,
l->convDesc,
l->dsrcTensorDesc,
CUDNN_CONVOLUTION_BWD_DATA_PREFER_FASTEST,
0,
&l->bd_algo);
cudnnGetConvolutionBackwardFilterAlgorithm(cudnn_handle(),
l->srcTensorDesc,
l->ddstTensorDesc,
l->convDesc,
l->dweightDesc,
CUDNN_CONVOLUTION_BWD_FILTER_PREFER_FASTEST,
0,
&l->bf_algo);
}
void backward_batchnorm_layer_gpu(const layer l, network_state state)
{
#ifdef CUDNN
float one = 1;
float zero = 0;
cudnnBatchNormalizationBackward(cudnn_handle(),
CUDNN_BATCHNORM_SPATIAL,
&one,
&zero,
&one,
&one,
l.dstTensorDesc,
l.x_gpu,
l.dstTensorDesc,
l.delta_gpu,
l.dstTensorDesc,
l.x_norm_gpu,
l.normTensorDesc,
l.scales_gpu,
l.scale_updates_gpu,
l.bias_updates_gpu,
.00001,
l.mean_gpu,
l.variance_gpu);
copy_ongpu(l.outputs*l.batch, l.x_norm_gpu, 1, l.delta_gpu, 1);
#else
backward_bias_gpu(l.bias_updates_gpu, l.delta_gpu, l.batch, l.out_c, l.out_w*l.out_h);
backward_scale_gpu(l.x_norm_gpu, l.delta_gpu, l.batch, l.out_c, l.out_w*l.out_h, l.scale_updates_gpu);
scale_bias_gpu(l.delta_gpu, l.scales_gpu, l.batch, l.out_c, l.out_h*l.out_w);
fast_mean_delta_gpu(l.delta_gpu, l.variance_gpu, l.batch, l.out_c, l.out_w*l.out_h, l.mean_delta_gpu);
fast_variance_delta_gpu(l.x_gpu, l.delta_gpu, l.mean_gpu, l.variance_gpu, l.batch, l.out_c, l.out_w*l.out_h, l.variance_delta_gpu);
normalize_delta_gpu(l.x_gpu, l.mean_gpu, l.variance_gpu, l.mean_delta_gpu, l.variance_delta_gpu, l.batch, l.out_c, l.out_w*l.out_h, l.delta_gpu);
#endif
if(l.type == BATCHNORM) copy_ongpu(l.outputs*l.batch, l.delta_gpu, 1, state.delta, 1);
}
void resize_convolutional_layer(convolutional_layer *l, int w, int h)
{
l->w = w;
l->h = h;
int out_w = convolutional_out_width(*l);
int out_h = convolutional_out_height(*l);
l->out_w = out_w;
l->out_h = out_h;
l->outputs = l->out_h * l->out_w * l->out_c;
l->inputs = l->w * l->h * l->c;
l->output = realloc(l->output,
l->batch*out_h * out_w * l->n*sizeof(float));
l->delta = realloc(l->delta,
l->batch*out_h * out_w * l->n*sizeof(float));
#ifdef GPU
cuda_free(l->delta_gpu);
cuda_free(l->output_gpu);
l->delta_gpu = cuda_make_array(l->delta, l->batch*out_h*out_w*l->n);
l->output_gpu = cuda_make_array(l->output, l->batch*out_h*out_w*l->n);
#ifdef CUDNN
cudnnSetTensor4dDescriptor(l->dsrcTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, l->batch, l->c, l->h, l->w);
cudnnSetTensor4dDescriptor(l->ddstTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, l->batch, l->out_c, l->out_h, l->out_w);
cudnnSetFilter4dDescriptor(l->dfilterDesc, CUDNN_DATA_FLOAT, CUDNN_TENSOR_NCHW, l->n, l->c, l->size, l->size);
cudnnSetTensor4dDescriptor(l->srcTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, l->batch, l->c, l->h, l->w);
cudnnSetTensor4dDescriptor(l->dstTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, l->batch, l->out_c, l->out_h, l->out_w);
cudnnSetFilter4dDescriptor(l->filterDesc, CUDNN_DATA_FLOAT, CUDNN_TENSOR_NCHW, l->n, l->c, l->size, l->size);
int padding = l->pad ? l->size/2 : 0;
cudnnSetConvolution2dDescriptor(l->convDesc, padding, padding, l->stride, l->stride, 1, 1, CUDNN_CROSS_CORRELATION);
cudnnGetConvolutionForwardAlgorithm(cudnn_handle(),
l->srcTensorDesc,
l->filterDesc,
l->convDesc,
l->dstTensorDesc,
CUDNN_CONVOLUTION_FWD_PREFER_FASTEST,
0,
&l->fw_algo);
cudnnGetConvolutionBackwardDataAlgorithm(cudnn_handle(),
l->filterDesc,
l->ddstTensorDesc,
l->convDesc,
l->dsrcTensorDesc,
CUDNN_CONVOLUTION_BWD_DATA_PREFER_FASTEST,
0,
&l->bd_algo);
cudnnGetConvolutionBackwardFilterAlgorithm(cudnn_handle(),
l->srcTensorDesc,
l->ddstTensorDesc,
l->convDesc,
l->dfilterDesc,
CUDNN_CONVOLUTION_BWD_FILTER_PREFER_FASTEST,
0,
&l->bf_algo);
#endif
#endif
l->workspace_size = get_workspace_size(*l);
}
convolutional_layer make_convolutional_layer(int batch, int h, int w, int c, int n, int size, int stride, int pad, ACTIVATION activation, int batch_normalize, int binary, int xnor)
{
int i;
convolutional_layer l = {0};
l.type = CONVOLUTIONAL;
l.h = h;
l.w = w;
l.c = c;
l.n = n;
l.binary = binary;
l.xnor = xnor;
l.batch = batch;
l.stride = stride;
l.size = size;
l.pad = pad;
l.batch_normalize = batch_normalize;
l.filters = calloc(c*n*size*size, sizeof(float));
l.filter_updates = calloc(c*n*size*size, sizeof(float));
l.biases = calloc(n, sizeof(float));
l.bias_updates = calloc(n, sizeof(float));
// float scale = 1./sqrt(size*size*c);
float scale = sqrt(2./(size*size*c));
for(i = 0; i < c*n*size*size; ++i) l.filters[i] = scale*rand_uniform(-1, 1);
int out_h = convolutional_out_height(l);
int out_w = convolutional_out_width(l);
l.out_h = out_h;
l.out_w = out_w;
l.out_c = n;
l.outputs = l.out_h * l.out_w * l.out_c;
l.inputs = l.w * l.h * l.c;
l.output = calloc(l.batch*out_h * out_w * n, sizeof(float));
l.delta = calloc(l.batch*out_h * out_w * n, sizeof(float));
if(binary){
l.binary_filters = calloc(c*n*size*size, sizeof(float));
l.cfilters = calloc(c*n*size*size, sizeof(char));
l.scales = calloc(n, sizeof(float));
}
if(xnor){
l.binary_filters = calloc(c*n*size*size, sizeof(float));
l.binary_input = calloc(l.inputs*l.batch, sizeof(float));
}
if(batch_normalize){
l.scales = calloc(n, sizeof(float));
l.scale_updates = calloc(n, sizeof(float));
for(i = 0; i < n; ++i){
l.scales[i] = 1;
}
l.mean = calloc(n, sizeof(float));
l.variance = calloc(n, sizeof(float));
l.rolling_mean = calloc(n, sizeof(float));
l.rolling_variance = calloc(n, sizeof(float));
}
#ifdef GPU
l.filters_gpu = cuda_make_array(l.filters, c*n*size*size);
l.filter_updates_gpu = cuda_make_array(l.filter_updates, c*n*size*size);
l.biases_gpu = cuda_make_array(l.biases, n);
l.bias_updates_gpu = cuda_make_array(l.bias_updates, n);
l.scales_gpu = cuda_make_array(l.scales, n);
l.scale_updates_gpu = cuda_make_array(l.scale_updates, n);
l.delta_gpu = cuda_make_array(l.delta, l.batch*out_h*out_w*n);
l.output_gpu = cuda_make_array(l.output, l.batch*out_h*out_w*n);
if(binary){
l.binary_filters_gpu = cuda_make_array(l.filters, c*n*size*size);
}
if(xnor){
l.binary_filters_gpu = cuda_make_array(l.filters, c*n*size*size);
l.binary_input_gpu = cuda_make_array(0, l.inputs*l.batch);
}
if(batch_normalize){
l.mean_gpu = cuda_make_array(l.mean, n);
l.variance_gpu = cuda_make_array(l.variance, n);
l.rolling_mean_gpu = cuda_make_array(l.mean, n);
l.rolling_variance_gpu = cuda_make_array(l.variance, n);
l.mean_delta_gpu = cuda_make_array(l.mean, n);
l.variance_delta_gpu = cuda_make_array(l.variance, n);
l.x_gpu = cuda_make_array(l.output, l.batch*out_h*out_w*n);
l.x_norm_gpu = cuda_make_array(l.output, l.batch*out_h*out_w*n);
}
#ifdef CUDNN
cudnnCreateTensorDescriptor(&l.srcTensorDesc);
cudnnCreateTensorDescriptor(&l.dstTensorDesc);
cudnnCreateFilterDescriptor(&l.filterDesc);
cudnnCreateTensorDescriptor(&l.dsrcTensorDesc);
cudnnCreateTensorDescriptor(&l.ddstTensorDesc);
cudnnCreateFilterDescriptor(&l.dfilterDesc);
cudnnCreateConvolutionDescriptor(&l.convDesc);
cudnnSetTensor4dDescriptor(l.dsrcTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, l.batch, l.c, l.h, l.w);
cudnnSetTensor4dDescriptor(l.ddstTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, l.batch, l.out_c, l.out_h, l.out_w);
cudnnSetFilter4dDescriptor(l.dfilterDesc, CUDNN_DATA_FLOAT, CUDNN_TENSOR_NCHW, l.n, l.c, l.size, l.size);
cudnnSetTensor4dDescriptor(l.srcTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, l.batch, l.c, l.h, l.w);
cudnnSetTensor4dDescriptor(l.dstTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, l.batch, l.out_c, l.out_h, l.out_w);
cudnnSetFilter4dDescriptor(l.filterDesc, CUDNN_DATA_FLOAT, CUDNN_TENSOR_NCHW, l.n, l.c, l.size, l.size);
int padding = l.pad ? l.size/2 : 0;
cudnnSetConvolution2dDescriptor(l.convDesc, padding, padding, l.stride, l.stride, 1, 1, CUDNN_CROSS_CORRELATION);
cudnnGetConvolutionForwardAlgorithm(cudnn_handle(),
l.srcTensorDesc,
l.filterDesc,
l.convDesc,
l.dstTensorDesc,
CUDNN_CONVOLUTION_FWD_PREFER_FASTEST,
0,
&l.fw_algo);
cudnnGetConvolutionBackwardDataAlgorithm(cudnn_handle(),
l.filterDesc,
l.ddstTensorDesc,
l.convDesc,
l.dsrcTensorDesc,
CUDNN_CONVOLUTION_BWD_DATA_PREFER_FASTEST,
0,
&l.bd_algo);
cudnnGetConvolutionBackwardFilterAlgorithm(cudnn_handle(),
l.srcTensorDesc,
l.ddstTensorDesc,
l.convDesc,
l.dfilterDesc,
CUDNN_CONVOLUTION_BWD_FILTER_PREFER_FASTEST,
0,
&l.bf_algo);
#endif
#endif
l.workspace_size = get_workspace_size(l);
l.activation = activation;
fprintf(stderr, "Convolutional Layer: %d x %d x %d image, %d filters -> %d x %d x %d image\n", h,w,c,n, out_h, out_w, n);
return l;
}