path: root/matrix/generic/cukernel.cu

#ifdef NERV_GENERIC_CUKERNEL
#include <assert.h>
#include <stdio.h>
#include "matrix.h"
#include "cuda.h"
#define CUDA_THREADS_N 16
#define CUDA_THREADS_NN (16 * 16)
#define CEIL_DIV(a, b) (((a) + (b) - 1) / (b))
__global__ void cudak_(sigmoid)(const MATRIX_ELEM *a, MATRIX_ELEM *b,
                        int nrow, int ncol, int stride) {
    int j = blockIdx.x * blockDim.x + threadIdx.x;
    int i = blockIdx.y * blockDim.y + threadIdx.y;
    long idx;
    if (i >= nrow || j >= ncol) return;
    idx = j + i * stride;
    b[idx] = 1.0 / (1.0 + exp(-a[idx]));
}

__global__ void cudak_(softmax_final)(const MATRIX_ELEM *a, MATRIX_ELEM *b,
                        const MATRIX_ELEM *max, const MATRIX_ELEM *deno,
                        int nrow, int ncol, int stride, int mstride) {
    int j = blockIdx.x * blockDim.x + threadIdx.x;
    int i = blockIdx.y * blockDim.y + threadIdx.y;
    long idx;
    if (i >= nrow || j >= ncol) return;
    idx = j + i * stride;
    b[idx] = exp(a[idx] - max[0 + i * mstride]) / deno[0 + i * mstride];
}

__global__ void cudak_(block_reduce_rowsum)(const MATRIX_ELEM *input,
                                MATRIX_ELEM *output,
                                const int istride, const int ostride,
                                const int n) {
    extern __shared__ MATRIX_ELEM cudak_(arr)[];
    int j = blockIdx.x * blockDim.x + threadIdx.x;
    cudak_(arr)[threadIdx.x] = j < n ? input[j + istride * blockIdx.y] : 0;
    __syncthreads();
    for (int offset = blockDim.x >> 1;  offset; offset >>= 1)
    {
        if (threadIdx.x < offset)
            cudak_(arr)[threadIdx.x] += cudak_(arr)[threadIdx.x + offset];
        __syncthreads();
    }
    if (threadIdx.x == 0)
        output[blockIdx.x + ostride * blockIdx.y] = cudak_(arr)[0];
}

__global__ void cudak_(block_reduce_colsum)(const MATRIX_ELEM *input,
                                MATRIX_ELEM *output,
                                const int istride, const int ostride,
                                const int n) {
    extern __shared__ MATRIX_ELEM cudak_(arr)[];
    int i = blockIdx.y * blockDim.y + threadIdx.y;
    cudak_(arr)[threadIdx.y] = i < n ? input[blockIdx.x + istride * i] : 0;
    __syncthreads();
    for (int offset = blockDim.y >> 1;  offset; offset >>= 1)
    {
        if (threadIdx.y < offset)
            cudak_(arr)[threadIdx.y] += cudak_(arr)[threadIdx.y + offset];
        __syncthreads();
    }
    if (threadIdx.y == 0)
        output[blockIdx.x + ostride * blockIdx.y] = cudak_(arr)[0];
}

__global__ void cudak_(block_reduce_softmax_rowsum)(const MATRIX_ELEM *input,
                                        MATRIX_ELEM *output,
                                        const MATRIX_ELEM *max,
                                        const int istride, const int ostride,
                                        const int mstride, const int n) {
    extern __shared__ MATRIX_ELEM cudak_(arr)[];
    int j = blockIdx.x * blockDim.x + threadIdx.x;
    cudak_(arr)[threadIdx.x] = j < n ? exp(input[j + istride * blockIdx.y] - \
                                    max[0 + mstride * blockIdx.y]) : 0;
    __syncthreads();
    for (int offset = blockDim.x >> 1;  offset; offset >>= 1)
    {
        if (threadIdx.x < offset)
            cudak_(arr)[threadIdx.x] += cudak_(arr)[threadIdx.x + offset];
        __syncthreads();
    }
    if (threadIdx.x == 0)
        output[blockIdx.x + ostride * blockIdx.y] = cudak_(arr)[0];
}

__global__ void cudak_(block_reduce_rowmax)(const MATRIX_ELEM *input,
                                MATRIX_ELEM *output,
                                const int istride, const int ostride,
                                const int n) {
    extern __shared__ MATRIX_ELEM cudak_(arr)[];
    int j = blockIdx.x * blockDim.x + threadIdx.x;
    cudak_(arr)[threadIdx.x] = j < n ? input[j + istride * blockIdx.y] : 0;
    __syncthreads();
    for (int offset = blockDim.x >> 1;  offset; offset >>= 1)
    {
        if (threadIdx.x < offset)
        {
            MATRIX_ELEM l = cudak_(arr)[threadIdx.x],
                  r = cudak_(arr)[threadIdx.x + offset];
            if (r > l) cudak_(arr)[threadIdx.x] = r;
        }
        __syncthreads();
    }
    if (threadIdx.x == 0)
        output[blockIdx.x + ostride * blockIdx.y] = cudak_(arr)[0];
}

__global__ void cudak_(add_row)(const MATRIX_ELEM *a, MATRIX_ELEM *b,
                                int nrow, int ncol, int