diff options
Diffstat (limited to 'nerv/lib/matrix/generic/cukernel.cu')
| -rw-r--r-- | nerv/lib/matrix/generic/cukernel.cu | 86 |
1 files changed, 85 insertions, 1 deletions
diff --git a/nerv/lib/matrix/generic/cukernel.cu b/nerv/lib/matrix/generic/cukernel.cu index d042d48..aa830b5 100644 --- a/nerv/lib/matrix/generic/cukernel.cu +++ b/nerv/lib/matrix/generic/cukernel.cu @@ -20,6 +20,19 @@ __global__ void cudak_(log_elem)(const MATRIX_ELEM *a, MATRIX_ELEM *b, b[idx] = log(tmp); } +__global__ void cudak_(thres_mask)(MATRIX_ELEM *a, MATRIX_ELEM *b, double thres, double low, double high, + 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; + if (b[idx] < thres) + a[idx] = low; + else + a[idx] = high; +} + __global__ void cudak_(mul_elem)(const MATRIX_ELEM *a, const MATRIX_ELEM *b, MATRIX_ELEM *c, int nrow, int ncol, int stride) { @@ -53,6 +66,29 @@ __global__ void cudak_(sigmoid_grad)(const MATRIX_ELEM *output, nerr[idx] = output[idx] * (1.0 - output[idx]) * err[idx]; } +__global__ void cudak_(tanh)(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] = (exp(a[idx]) - exp(-a[idx])) / (exp(a[idx]) + exp(-a[idx])); //could cause nan + b[idx] = tanh(a[idx]); +} + +__global__ void cudak_(tanh_grad)(const MATRIX_ELEM *output, + const MATRIX_ELEM *err, + MATRIX_ELEM *nerr, + 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; + nerr[idx] = (1.0 - output[idx] * output[idx]) * err[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) { @@ -225,14 +261,18 @@ __global__ void cudak_(clip)(MATRIX_ELEM *a, a[j + i * stride] = val_1; } +#ifdef __NERV_FUTURE_CUDA_7 __global__ void cudak_(update_select_rows)(MATRIX_ELEM *c, const MATRIX_ELEM *a, const MATRIX_ELEM *idx, int nrow_a, int ncol_a, int stride_c, int stride_a, double alpha, double beta) { int j = blockIdx.x * blockDim.x + threadIdx.x; int i = blockIdx.y * blockDim.y + threadIdx.y; if (i >= nrow_a || j >= ncol_a) return; int i_c = lrintf(idx[i]); - c[j + i_c * stride_c] = c[j + i_c * stride_c] * (1 - beta * alpha) + a[j + i * stride_a] * alpha; + //critical: i_c could conflict among threads(same index in the idx array), so atomicAdd is used + //c[j + i_c * stride_c] = c[j + i_c * stride_c] * (1 - beta * alpha) + a[j + i * stride_a] * alpha; + atomicAdd_nvidia(c + j + i_c * stride_c, c[j + i_c * stride_c] * (- beta * alpha) + a[j + i * stride_a] * alpha); } +#endif __global__ void cudak_(expand_frm)(const MATRIX_ELEM *a, MATRIX_ELEM *b, int nrow, int ncol, @@ -349,6 +389,48 @@ extern "C" { cudaStreamSynchronize(0); } + void cudak_(cuda_rand_uniform)(const Matrix *a) { + #ifdef MATRIX_USE_FLOAT + curandGenerateUniform(*(a->curand_gen), MATRIX_ELEM_PTR(a), a->nrow * a->stride / sizeof(MATRIX_ELEM)); + #endif + #ifdef MATRIX_USE_DOUBLE + curandGenerateUniformDouble(*(a->curand_gen), MATRIX_ELEM_PTR(a), a->nrow * a->stride / sizeof(MATRIX_ELEM)); + #endif + } + + void cudak_(cuda_thres_mask)(const Matrix *a, const Matrix *b, double thres, double low, double high) { + dim3 threadsPerBlock(CUDA_THREADS_N, CUDA_THREADS_N); + dim3 numBlocks(CEIL_DIV(a->ncol, threadsPerBlock.x), + CEIL_DIV(a->nrow, threadsPerBlock.y)); + cudak_(thres_mask)<<<numBlocks, threadsPerBlock>>> \ + (MATRIX_ELEM_PTR(a), MATRIX_ELEM_PTR(b), + thres, low, high, a->nrow, a->ncol, a->stride / sizeof(MATRIX_ELEM)); + cudaStreamSynchronize(0); + } + + void cudak_(cuda_tanh)(const Matrix *a, Matrix *b) { + dim3 threadsPerBlock(CUDA_THREADS_N, CUDA_THREADS_N); + dim3 numBlocks(CEIL_DIV(b->ncol, threadsPerBlock.x), + CEIL_DIV(b->nrow, threadsPerBlock.y)); + cudak_(tanh)<<<numBlocks, threadsPerBlock>>> \ + (MATRIX_ELEM_PTR(a), MATRIX_ELEM_PTR(b), b->nrow, b->ncol, + b->stride / sizeof(MATRIX_ELEM)); + cudaStreamSynchronize(0); + } + + void cudak_(cuda_tanh_grad)(const Matrix *output, + const Matrix *err, Matrix *nerr) { + dim3 threadsPerBlock(CUDA_THREADS_N, CUDA_THREADS_N); + dim3 numBlocks(CEIL_DIV(nerr->ncol, threadsPerBlock.x), + CEIL_DIV(nerr->nrow, threadsPerBlock.y)); + cudak_(tanh_grad)<<<numBlocks, threadsPerBlock>>> \ + (MATRIX_ELEM_PTR(output), MATRIX_ELEM_PTR(err), + MATRIX_ELEM_PTR(nerr), + nerr->nrow, nerr->ncol, + nerr->stride / sizeof(MATRIX_ELEM)); + cudaStreamSynchronize(0); + } + void cudak_(cuda_rowsum)(const Matrix *a, Matrix *b) { dim3 block(CUDA_THREADS_NN, 1); int ncol = a->ncol; @@ -550,6 +632,7 @@ extern "C" { cudaStreamSynchronize(0); } +#ifdef __NERV_FUTURE_CUDA_7 void cudak_(cuda_update_select_rows)(Matrix *c, const Matrix *a, const Matrix *idx, double alpha, double beta) { dim3 threadsPerBlock(CUDA_THREADS_N, CUDA_THREADS_N); dim3 numBlocks(CEIL_DIV(a->ncol, threadsPerBlock.x), @@ -560,6 +643,7 @@ extern "C" { a->stride / sizeof(MATRIX_ELEM), alpha, beta); cudaStreamSynchronize(0); } +#endif void cudak_(cuda_expand_frm)(const Matrix *a, Matrix *b, int context) { dim3 threadsPerBlock(CUDA_THREADS_N, CUDA_THREADS_N); |