add sigmoid for cumatrix

5 files changed, 40 insertions, 1 deletions

diff --git a/Makefile b/Makefile
index 4514263..bfda339 100644
--- a/Makefile
+++ b/Makefile
@@ -1,5 +1,5 @@
 .PHONY: all clean luajit
-OBJS := oop_example.o nerv.o luaT.o common.o matrix/matrix.o matrix/cumatrix.o matrix/init.o
+OBJS := oop_example.o nerv.o luaT.o common.o matrix/matrix.o matrix/cumatrix.o matrix/init.o matrix/cukernel.o
 LIBS := libnerv.so
 LUA_LIBS := matrix/init.lua nerv.lua
 INCLUDE := -I build/luajit-2.0/include/luajit-2.0/ -DLUA_USE_APICHECK
@@ -10,6 +10,7 @@ LDFLAGS := -L$(CUDA_BASE)/lib64/  -Wl,-rpath=$(CUDA_BASE)/lib64/ -lcudart -lcubl
 CFLAGS :=
 OBJ_DIR := build/objs
 LUA_DIR := build/lua
+NVCC := $(CUDA_BASE)/bin/nvcc
 
 OBJS := $(addprefix $(OBJ_DIR)/,$(OBJS))
 LIBS := $(addprefix $(OBJ_DIR)/,$(LIBS))
@@ -28,6 +29,8 @@ $(OBJ_DIR)/%.o: %.c
 	gcc -c -o $@ $< $(INCLUDE) -fPIC $(CFLAGS)
 $(OBJ_DIR)/matrix/%.o: matrix/%.c
 	gcc -c -o $@ $< $(INCLUDE) -fPIC $(CFLAGS)
+$(OBJ_DIR)/matrix/cukernel.o: matrix/cukernel.cu
+	$(NVCC) -c -o $@ $< -Xcompiler -fPIC $(INCLUDE) $(CFLAGS)
 $(LUA_DIR)/%.lua: %.lua
 	cp $< $@
 $(OBJ_DIR)/luaT.o:
diff --git a/cumatrix_example.lua b/cumatrix_example.lua
index ccd88b8..ce11eea 100644
--- a/cumatrix_example.lua
+++ b/cumatrix_example.lua
@@ -25,3 +25,4 @@ d[1][1] = 2
 d[2][2] = 3
 print(d)
 print(t * d)
+print(t:sigmoid())
diff --git a/matrix/cukernel.cu b/matrix/cukernel.cu
new file mode 100644
index 0000000..91e7e35
--- /dev/null
+++ b/matrix/cukernel.cu
@@ -0,0 +1,21 @@
+#include "generic/matrix.h"
+#define CUDA_THREADS_N 16
+#define CEIL_DIV(a, b) (((a) + (b) - 1) / (b))
+__global__ void sigmoid(const float *a, float *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]));
+}
+
+extern "C" void cuda_sigmoid(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));
+    sigmoid<<<numBlocks, threadsPerBlock>>>(a->data.f, b->data.f, b->nrow, b->ncol,
+                                            b->stride / sizeof(float));
+}
diff --git a/matrix/cukernel.h b/matrix/cukernel.h
new file mode 100644
index 0000000..5b9e3a6
--- /dev/null
+++ b/matrix/cukernel.h
@@ -0,0 +1,4 @@
+#ifndef NERV_CUKERNEL_H
+#define NERV_CUKERNEL_H
+void cuda_sigmoid(const Matrix *a, Matrix *b);
+#endif
diff --git a/matrix/cumatrix.c b/matrix/cumatrix.c
index 9c2878a..7759ca1 100644
--- a/matrix/cumatrix.c
+++ b/matrix/cumatrix.c
@@ -7,6 +7,7 @@
 #define nerv_float_matrix_(NAME) nerv_float_matrix_cuda_ ## NAME
 #include "../common.h"
 #include "generic/matrix.h"
+#include "cukernel.h"
 #include "cuda.h"
 #include "driver_types.h"
 #include "cublas_v2.h"
@@ -56,9 +57,18 @@ static int nerv_float_matrix_(mul)(lua_State *L) {
     return 1;
 }
 
+static int nerv_float_matrix_(sigmoid)(lua_State *L) {
+    Matrix *a = luaT_checkudata(L, 1, nerv_float_matrix_(tname));
+    Matrix *b = nerv_float_matrix_(new_)(a->nrow, a->ncol);
+    cuda_sigmoid(a, b);
+    luaT_pushudata(L, b, nerv_float_matrix_(tname));
+    return 1;
+}
+
 static const luaL_Reg nerv_float_matrix_(extra_methods)[] = {
     {"__add__", nerv_float_matrix_(add)},
     {"__mul__", nerv_float_matrix_(mul)},
+    {"sigmoid", nerv_float_matrix_(sigmoid)},
     {NULL, NULL}
 };