From e4dedc2992149d245ea65132131253072d3276b8 Mon Sep 17 00:00:00 2001
From: Determinant <ted.sybil@gmail.com>
Date: Thu, 28 May 2015 16:14:21 +0800
Subject: add mul_elem and log_elem; add dimension checking in cumatrix methods
---
matrix/generic/cukernel.cu | 43 +++++++++++++++++++++++++++++++++++++++++++
matrix/generic/cumatrix.c | 27 ++++++++++++++++++++++++---
2 files changed, 67 insertions(+), 3 deletions(-)
(limited to 'matrix/generic')
diff --git a/matrix/generic/cukernel.cu b/matrix/generic/cukernel.cu
index 517393e..0e3d3cf 100644
--- a/matrix/generic/cukernel.cu
+++ b/matrix/generic/cukernel.cu
@@ -6,6 +6,27 @@
#define CUDA_THREADS_N 16
#define CUDA_THREADS_NN (16 * 16)
#define CEIL_DIV(a, b) (((a) + (b) - 1) / (b))
+__global__ void cudak_(log_elem)(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] = log(a[idx]);
+}
+
+__global__ void cudak_(mul_elem)(const MATRIX_ELEM *a, const MATRIX_ELEM *b,
+ MATRIX_ELEM *c,
+ 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;
+ c[idx] = a[idx] * b[idx];
+}
+
__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;
@@ -136,6 +157,28 @@ __global__ void cudak_(fill)(MATRIX_ELEM *a,
extern "C" {
#include "../cukernel.h"
+ void cudak_(cuda_log_elem)(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_(log_elem)<<<numBlocks, threadsPerBlock>>> \
+ (MATRIX_ELEM_PTR(a), MATRIX_ELEM_PTR(b),
+ b->nrow, b->ncol, b->stride / sizeof(MATRIX_ELEM));
+ }
+
+ void cudak_(cuda_mul_elem)(const Matrix *a, const Matrix *b,
+ Matrix *c) {
+ dim3 threadsPerBlock(CUDA_THREADS_N,
+ CUDA_THREADS_N);
+ dim3 numBlocks(CEIL_DIV(b->ncol, threadsPerBlock.x),
+ CEIL_DIV(b->nrow, threadsPerBlock.y));
+ cudak_(mul_elem)<<<numBlocks, threadsPerBlock>>> \
+ (MATRIX_ELEM_PTR(a), MATRIX_ELEM_PTR(b),
+ MATRIX_ELEM_PTR(c),
+ b->nrow, b->ncol, b->stride / sizeof(MATRIX_ELEM));
+ }
+
void cudak_(cuda_sigmoid)(const Matrix *a, Matrix *b) {
dim3 threadsPerBlock(CUDA_THREADS_N,
CUDA_THREADS_N);
diff --git a/matrix/generic/cumatrix.c b/matrix/generic/cumatrix.c
index 2deb7a3..ed64bbf 100644
--- a/matrix/generic/cumatrix.c
+++ b/matrix/generic/cumatrix.c
@@ -48,6 +48,7 @@ static int nerv_matrix_(add)(lua_State *L) {
MATRIX_ELEM alpha = luaL_checknumber(L, 4); /* alpha */
MATRIX_ELEM beta = luaL_checknumber(L, 5); /* alpha */
CHECK_SAME_DIMENSION(a, b);
+ CHECK_SAME_DIMENSION(a, c);
nerv_matrix_(add_)(a, b, c, alpha, beta);
return 0;
}
@@ -118,6 +119,7 @@ static int nerv_matrix_(softmax)(lua_State *L) {
Matrix *b = luaT_checkudata(L, 1, nerv_matrix_(tname));
Matrix *max = nerv_matrix_(new_)(a->nrow, 1);
Matrix *dno = nerv_matrix_(new_)(a->nrow, 1);
+ CHECK_SAME_DIMENSION(a, b);
cudak_(cuda_rowmax)(a, max);
cudak_(cuda_softmax_denominator)(a, max, dno);
cudak_(cuda_softmax_final)(a, max, dno, b);
@@ -230,25 +232,44 @@ static int nerv_matrix_(trans)(lua_State *L) {
return 1;
}
+static int nerv_matrix_(mul_elem)(lua_State *L) {
+ Matrix *a = luaT_checkudata(L, 2, nerv_matrix_(tname));
+ Matrix *b = luaT_checkudata(L, 3, nerv_matrix_(tname));
+ Matrix *c = luaT_checkudata(L, 1, nerv_matrix_(tname));
+ CHECK_SAME_DIMENSION(a, b);
+ CHECK_SAME_DIMENSION(a, c);
+ cudak_(cuda_mul_elem)(a, b, c);
+ return 0;
+}
+
+static int nerv_matrix_(log_elem)(lua_State *L) {
+ Matrix *a = luaT_checkudata(L, 2, nerv_matrix_(tname));
+ Matrix *b = luaT_checkudata(L, 1, nerv_matrix_(tname));
+ CHECK_SAME_DIMENSION(a, b);
+ cudak_(cuda_log_elem)(a, b);
+ return 0;
+}
static const luaL_Reg nerv_matrix_(extra_methods)[] = {
{"create", nerv_matrix_(create)},
- {"softmax", nerv_matrix_(softmax)},
{"colsum", nerv_matrix_(colsum)},
{"rowsum", nerv_matrix_(rowsum)},
{"rowmax", nerv_matrix_(rowmax)},
+ {"trans", nerv_matrix_(trans)},
+ /* in-place calc */
{"copy_fromh", nerv_matrix_(copy_fromh)},
{"copy_fromd", nerv_matrix_(copy_fromd)},
{"copy_toh", nerv_matrix_(copy_toh)},
{"copy_tod", nerv_matrix_(copy_tod)},
- {"trans", nerv_matrix_(trans)},
- /* in-place calc */
{"add", nerv_matrix_(add)},
{"mul", nerv_matrix_(mul)},
{"add_row", nerv_matrix_(add_row)},
{"fill", nerv_matrix_(fill)},
{"sigmoid", nerv_matrix_(sigmoid)},
{"sigmoid_grad", nerv_matrix_(sigmoid_grad)},
+ {"softmax", nerv_matrix_(softmax)},
+ {"mul_elem", nerv_matrix_(mul_elem)},
+ {"log_elem", nerv_matrix_(log_elem)},
{NULL, NULL}
};
--
cgit v1.2.3-70-g09d2