| | |
| | | dim3 dimGrid((size-1)/BLOCK + 1, n, batch); |
| | | dim3 dimBlock(BLOCK, 1, 1); |
| | | |
| | | scale_bias_kernel<<<dimGrid, dimBlock>>>(output, biases, n, size); |
| | | scale_bias_kernel<<<dimGrid, dimBlock, 0, get_cuda_stream()>>>(output, biases, n, size); |
| | | check_error(cudaPeekAtLastError()); |
| | | } |
| | | |
| | |
| | | dim3 dimGrid((size-1)/BLOCK + 1, n, batch); |
| | | dim3 dimBlock(BLOCK, 1, 1); |
| | | |
| | | add_bias_kernel<<<dimGrid, dimBlock>>>(output, biases, n, size); |
| | | add_bias_kernel<<<dimGrid, dimBlock, 0, get_cuda_stream()>>>(output, biases, n, size); |
| | | check_error(cudaPeekAtLastError()); |
| | | } |
| | | |
| | |
| | | } |
| | | |
| | | |
| | | __global__ void adam_kernel(int N, float *x, float *m, float *v, float B1, float B2, float rate, float eps, int t) |
| | | { |
| | | int index = (blockIdx.x + blockIdx.y*gridDim.x) * blockDim.x + threadIdx.x; |
| | | if (index >= N) return; |
| | | |
| | | x[index] = x[index] - (rate * sqrt(1.-pow(B2, t)) / (1.-pow(B1, t)) * m[index] / (sqrt(v[index]) + eps)); |
| | | //if(index == 0) printf("%f %f %f %f\n", m[index], v[index], (rate * sqrt(1.-pow(B2, t)) / (1.-pow(B1, t)) * m[index] / (sqrt(v[index]) + eps))); |
| | | } |
| | | |
| | | extern "C" void adam_gpu(int n, float *x, float *m, float *v, float B1, float B2, float rate, float eps, int t) |
| | | { |
| | | adam_kernel<<<cuda_gridsize(n), BLOCK>>>(n, x, m, v, B1, B2, rate, eps, t); |
| | | check_error(cudaPeekAtLastError()); |
| | | } |
| | | |
| | | __global__ void normalize_kernel(int N, float *x, float *mean, float *variance, int batch, int filters, int spatial) |
| | | { |
| | | int index = (blockIdx.x + blockIdx.y*gridDim.x) * blockDim.x + threadIdx.x; |
| | |
| | | local[id] += (i+id < spatial) ? delta[index] : 0; |
| | | } |
| | | } |
| | | __syncthreads(); |
| | | |
| | | if(id == 0){ |
| | | mean_delta[filter] = 0; |
| | |
| | | local[id] += (i+id < spatial) ? delta[index]*(x[index] - mean[filter]) : 0; |
| | | } |
| | | } |
| | | __syncthreads(); |
| | | |
| | | if(id == 0){ |
| | | variance_delta[filter] = 0; |
| | |
| | | extern "C" void normalize_gpu(float *x, float *mean, float *variance, int batch, int filters, int spatial) |
| | | { |
| | | size_t N = batch*filters*spatial; |
| | | normalize_kernel<<<cuda_gridsize(N), BLOCK>>>(N, x, mean, variance, batch, filters, spatial); |
| | | normalize_kernel<<<cuda_gridsize(N), BLOCK, 0, get_cuda_stream()>>>(N, x, mean, variance, batch, filters, spatial); |
| | | check_error(cudaPeekAtLastError()); |
| | | } |
| | | |
| | |
| | | local[id] += (i+id < spatial) ? x[index] : 0; |
| | | } |
| | | } |
| | | __syncthreads(); |
| | | |
| | | if(id == 0){ |
| | | mean[filter] = 0; |
| | |
| | | local[id] += (i+id < spatial) ? pow((x[index] - mean[filter]), 2) : 0; |
| | | } |
| | | } |
| | | __syncthreads(); |
| | | |
| | | if(id == 0){ |
| | | variance[filter] = 0; |
| | |
| | | |
| | | extern "C" void fast_mean_gpu(float *x, int batch, int filters, int spatial, float *mean) |
| | | { |
| | | fast_mean_kernel<<<filters, BLOCK>>>(x, batch, filters, spatial, mean); |
| | | fast_mean_kernel<<<filters, BLOCK, 0, get_cuda_stream()>>>(x, batch, filters, spatial, mean); |
| | | check_error(cudaPeekAtLastError()); |
| | | } |
| | | |
| | | extern "C" void fast_variance_gpu(float *x, float *mean, int batch, int filters, int spatial, float *variance) |
| | | { |
| | | fast_variance_kernel<<<filters, BLOCK>>>(x, mean, batch, filters, spatial, variance); |
| | | fast_variance_kernel<<<filters, BLOCK, 0, get_cuda_stream() >>>(x, mean, batch, filters, spatial, variance); |
| | | check_error(cudaPeekAtLastError()); |
| | | } |
| | | |
| | |
| | | |
| | | extern "C" void pow_ongpu(int N, float ALPHA, float * X, int INCX, float * Y, int INCY) |
| | | { |
| | | pow_kernel<<<cuda_gridsize(N), BLOCK>>>(N, ALPHA, X, INCX, Y, INCY); |
| | | pow_kernel<<<cuda_gridsize(N), BLOCK, 0, get_cuda_stream() >>>(N, ALPHA, X, INCX, Y, INCY); |
| | | check_error(cudaPeekAtLastError()); |
| | | } |
| | | |
| | | extern "C" void axpy_ongpu_offset(int N, float ALPHA, float * X, int OFFX, int INCX, float * Y, int OFFY, int INCY) |
| | | { |
| | | axpy_kernel<<<cuda_gridsize(N), BLOCK>>>(N, ALPHA, X, OFFX, INCX, Y, OFFY, INCY); |
| | | axpy_kernel<<<cuda_gridsize(N), BLOCK, 0, get_cuda_stream()>>>(N, ALPHA, X, OFFX, INCX, Y, OFFY, INCY); |
| | | check_error(cudaPeekAtLastError()); |
| | | } |
| | | |
| | |
| | | |
| | | extern "C" void copy_ongpu_offset(int N, float * X, int OFFX, int INCX, float * Y, int OFFY, int INCY) |
| | | { |
| | | copy_kernel<<<cuda_gridsize(N), BLOCK>>>(N, X, OFFX, INCX, Y, OFFY, INCY); |
| | | copy_kernel<<<cuda_gridsize(N), BLOCK, 0, get_cuda_stream()>>>(N, X, OFFX, INCX, Y, OFFY, INCY); |
| | | check_error(cudaPeekAtLastError()); |
| | | } |
| | | |
| | | __global__ void flatten_kernel(int N, float *x, int spatial, int layers, int batch, int forward, float *out) |
| | | { |
| | | int i = (blockIdx.x + blockIdx.y*gridDim.x) * blockDim.x + threadIdx.x; |
| | | if(i >= N) return; |
| | | int in_s = i%spatial; |
| | | i = i/spatial; |
| | | int in_c = i%layers; |
| | | i = i/layers; |
| | | int b = i; |
| | | |
| | | int i1 = b*layers*spatial + in_c*spatial + in_s; |
| | | int i2 = b*layers*spatial + in_s*layers + in_c; |
| | | |
| | | if (forward) out[i2] = x[i1]; |
| | | else out[i1] = x[i2]; |
| | | } |
| | | |
| | | extern "C" void flatten_ongpu(float *x, int spatial, int layers, int batch, int forward, float *out) |
| | | { |
| | | int size = spatial*batch*layers; |
| | | flatten_kernel<<<cuda_gridsize(size), BLOCK, 0, get_cuda_stream()>>>(size, x, spatial, layers, batch, forward, out); |
| | | check_error(cudaPeekAtLastError()); |
| | | } |
| | | |
| | | extern "C" void reorg_ongpu(float *x, int w, int h, int c, int batch, int stride, int forward, float *out) |
| | | { |
| | | int size = w*h*c*batch; |
| | | reorg_kernel<<<cuda_gridsize(size), BLOCK>>>(size, x, w, h, c, batch, stride, forward, out); |
| | | reorg_kernel<<<cuda_gridsize(size), BLOCK, 0, get_cuda_stream()>>>(size, x, w, h, c, batch, stride, forward, out); |
| | | check_error(cudaPeekAtLastError()); |
| | | } |
| | | |
| | | extern "C" void mask_ongpu(int N, float * X, float mask_num, float * mask) |
| | | { |
| | | mask_kernel<<<cuda_gridsize(N), BLOCK>>>(N, X, mask_num, mask); |
| | | mask_kernel<<<cuda_gridsize(N), BLOCK, 0, get_cuda_stream() >>>(N, X, mask_num, mask); |
| | | check_error(cudaPeekAtLastError()); |
| | | } |
| | | |
| | |
| | | |
| | | extern "C" void scal_ongpu(int N, float ALPHA, float * X, int INCX) |
| | | { |
| | | scal_kernel<<<cuda_gridsize(N), BLOCK>>>(N, ALPHA, X, INCX); |
| | | scal_kernel<<<cuda_gridsize(N), BLOCK, 0, get_cuda_stream()>>>(N, ALPHA, X, INCX); |
| | | check_error(cudaPeekAtLastError()); |
| | | } |
| | | |
| | |
| | | |
| | | extern "C" void fill_ongpu(int N, float ALPHA, float * X, int INCX) |
| | | { |
| | | fill_kernel<<<cuda_gridsize(N), BLOCK>>>(N, ALPHA, X, INCX); |
| | | fill_kernel<<<cuda_gridsize(N), BLOCK, 0, get_cuda_stream()>>>(N, ALPHA, X, INCX); |
| | | check_error(cudaPeekAtLastError()); |
| | | } |
| | | |
| | |
| | | } |
| | | |
| | | |
| | | __global__ void softmax_kernel(int n, int batch, float *input, float temp, float *output) |
| | | __device__ void softmax_device(int n, float *input, float temp, float *output) |
| | | { |
| | | int b = (blockIdx.x + blockIdx.y*gridDim.x) * blockDim.x + threadIdx.x; |
| | | if(b >= batch) return; |
| | | |
| | | int i; |
| | | float sum = 0; |
| | | float largest = -INFINITY; |
| | | for(i = 0; i < n; ++i){ |
| | | int val = input[i+b*n]; |
| | | int val = input[i]; |
| | | largest = (val>largest) ? val : largest; |
| | | } |
| | | for(i = 0; i < n; ++i){ |
| | | sum += exp(input[i+b*n]/temp-largest/temp); |
| | | float e = exp(input[i]/temp - largest/temp); |
| | | sum += e; |
| | | output[i] = e; |
| | | } |
| | | sum = (sum != 0) ? largest/temp+log(sum) : largest-100; |
| | | for(i = 0; i < n; ++i){ |
| | | output[i+b*n] = exp(input[i+b*n]/temp-sum); |
| | | output[i] /= sum; |
| | | } |
| | | } |
| | | |
| | | extern "C" void softmax_gpu(float *input, int n, int groups, float temp, float *output, cudaStream_t stream) |
| | | __global__ void softmax_kernel(int n, int offset, int batch, float *input, float temp, float *output) |
| | | { |
| | | int b = (blockIdx.x + blockIdx.y*gridDim.x) * blockDim.x + threadIdx.x; |
| | | if(b >= batch) return; |
| | | softmax_device(n, input + b*offset, temp, output + b*offset); |
| | | } |
| | | |
| | | extern "C" void softmax_gpu(float *input, int n, int offset, int groups, float temp, float *output) |
| | | { |
| | | int inputs = n; |
| | | int batch = groups; |
| | | softmax_kernel<<<cuda_gridsize(batch), BLOCK, 0, stream>>>(inputs, batch, input, temp, output); |
| | | softmax_kernel<<<cuda_gridsize(batch), BLOCK, 0, get_cuda_stream()>>>(inputs, offset, batch, input, temp, output); |
| | | check_error(cudaPeekAtLastError()); |
| | | } |