add profiling; add ce accurarcy; several other changes

16 files changed, 416 insertions, 38 deletions

diff --git a/Makefile b/Makefile
index f0d319f..5c6fa7b 100644
--- a/Makefile
+++ b/Makefile
@@ -16,7 +16,7 @@ CUDA_BASE := /usr/local/cuda-6.5
 CUDA_INCLUDE := -I $(CUDA_BASE)/include/
 INCLUDE += $(CUDA_INCLUDE)
 LDFLAGS := -L$(CUDA_BASE)/lib64/  -Wl,-rpath=$(CUDA_BASE)/lib64/ -lcudart -lcublas
-CFLAGS := -Wall -Wextra
+CFLAGS := -Wall -Wextra -O2
 OBJ_DIR := $(BUILD_DIR)/objs
 LUA_DIR := $(BUILD_DIR)/lua
 LIB_DIR := $(BUILD_DIR)/lib
diff --git a/common.c b/common.c
index c60c1ec..355d7ff 100644
--- a/common.c
+++ b/common.c
@@ -1,5 +1,3 @@
-#ifndef NERV_COMMON_H
-#define NERV_COMMON_H
 #include "common.h"
 #include <stdarg.h>
 int nerv_error(lua_State *L, const char *err_mesg_fmt, ...) {
@@ -24,4 +22,55 @@ void luaN_append_methods(lua_State *L, const luaL_Reg *mlist) {
         lua_setfield(L, -2, mlist->name);
     }
 }
-#endif
+
+HashMap *hashmap_create(size_t size, HashKey_t hfunc, HashMapCmp_t cmp) {
+    HashMap *res = (HashMap *)malloc(sizeof(HashMap));
+    res->bucket = calloc(size, sizeof(HashNode));
+    res->cmp = cmp;
+    res->hfunc = hfunc;
+    res->size = size;
+    return res;
+}
+
+void *hashmap_getval(HashMap *h, const char *key) {
+    size_t idx = h->hfunc(key) % h->size;
+    HashNode *ptr;
+    for (ptr = h->bucket[idx]; ptr; ptr = ptr->next)
+    {
+        if (!h->cmp(ptr->key, key))
+            return ptr->val;
+    }
+    return NULL;
+}
+
+void hashmap_setval(HashMap *h, const char *key, void *val) {
+    size_t idx = h->hfunc(key) % h->size;
+    HashNode *ptr = malloc(sizeof(HashNode));
+    ptr->next = h->bucket[idx];
+    h->bucket[idx] = ptr;
+    ptr->key = key;
+    ptr->val = val;
+}
+
+void hashmap_clear(HashMap *h) {
+    size_t i;
+    for (i = 0; i < h->size; i++)
+    {
+        HashNode *ptr, *nptr;
+        for (ptr = h->bucket[i]; ptr; ptr = nptr)
+        {
+            nptr = ptr->next;
+            free(ptr->val);
+            free(ptr);
+        }
+        h->bucket[i] = NULL;
+    }
+}
+
+size_t bkdr_hash(const char *key) {
+    unsigned int seed = 131;
+    unsigned int res = 0;
+    while (*key)
+        res = res * seed + *key++;
+    return res;
+}
diff --git a/common.h b/common.h
index 51e90ee..8be19b0 100644
--- a/common.h
+++ b/common.h
@@ -1,3 +1,5 @@
+#ifndef NERV_COMMON_H
+#define NERV_COMMON_H
 #include "lua.h"
 #include "lauxlib.h"
 #include "lualib.h"
@@ -5,6 +7,29 @@
 #include <stdio.h>
 #include <stdlib.h>
 
+typedef struct HashNode {
+    const char *key;
+    void *val;
+    struct HashNode *next;
+} HashNode;
+
+typedef int (*HashMapCmp_t)(const char *a, const char *b);
+typedef size_t (*HashKey_t)(const char *key);
+
+typedef struct HashMap {
+    HashNode **bucket;
+    HashMapCmp_t cmp;
+    HashKey_t hfunc;
+    size_t size;
+} HashMap;
+
+HashMap *hashmap_create(size_t size, HashKey_t hfunc, HashMapCmp_t cmp);
+void *hashmap_getval(HashMap *h, const char *key);
+void hashmap_setval(HashMap *h, const char *key, void *val);
+
+size_t bkdr_hash(const char *key);
+
 int nerv_error(lua_State *L, const char *err_mesg_fmt, ...); 
 int nerv_error_method_not_implemented(lua_State *L);
 void luaN_append_methods(lua_State *L, const luaL_Reg *mlist);
+#endif
diff --git a/examples/test_nn_lib.lua b/examples/test_nn_lib.lua
index 9600917..04fd7d6 100644
--- a/examples/test_nn_lib.lua
+++ b/examples/test_nn_lib.lua
@@ -116,7 +116,8 @@ tnet_reader = nerv.TNetReader(gconf,
 
 buffer = nerv.SGDBuffer(gconf,
     {
-        buffer_size = 8192,
+        buffer_size = 81920,
+        -- randomize = true,
         readers = {
             { reader = tnet_reader,
               data = {main_scp = 429, ref = 1}}
@@ -126,10 +127,11 @@ buffer = nerv.SGDBuffer(gconf,
 sm = sublayer_repo:get_layer("softmax_ce0")
 main = layer_repo:get_layer("main")
 main:init(gconf.batch_size)
-cnt = 0
+gconf.cnt = 0
 for data in buffer.get_data, buffer do
-    if cnt == 1000 then break end
-    cnt = cnt + 1
+    if gconf.cnt == 1000 then break end
+    gconf.cnt = gconf.cnt + 1
+
     input = {data.main_scp, data.ref}
     output = {}
     err_input = {}
@@ -140,7 +142,10 @@ for data in buffer.get_data, buffer do
     main:update(err_input, input, output)
 
     nerv.utils.printf("cross entropy: %.8f\n", sm.total_ce)
-    nerv.utils.printf("frames: %.8f\n", sm.total_frames)
+    nerv.utils.printf("correct: %d\n", sm.total_correct)
+    nerv.utils.printf("frames: %d\n", sm.total_frames)
     nerv.utils.printf("err/frm: %.8f\n", sm.total_ce / sm.total_frames)
+    nerv.utils.printf("accuracy: %.8f\n", sm.total_correct / sm.total_frames)
     collectgarbage("collect")
 end
+nerv.Matrix.print_profile()
diff --git a/io/chunk_file.c b/io/chunk_file.c
index ce346c5..4e987b7 100644
--- a/io/chunk_file.c
+++ b/io/chunk_file.c
@@ -268,7 +268,7 @@ int nerv_chunk_file_handle_destroy(lua_State *L) {
     return 0;
 }
 
-static int nerv_chunk_destroy(lua_State *L) {
+static int nerv_chunk_info_destroy(lua_State *L) {
     ChunkInfo *pci = luaT_checkudata(L, 1, nerv_chunk_info_tname);
     free(pci);
     return 0;
@@ -298,7 +298,7 @@ void nerv_chunk_file_init(lua_State *L) {
     luaT_newmetatable(L, nerv_chunk_file_handle_tname, NULL,
                         NULL, nerv_chunk_file_handle_destroy, NULL);
     luaT_newmetatable(L, nerv_chunk_info_tname, NULL,
-                        NULL, nerv_chunk_destroy, NULL);
+                        NULL, nerv_chunk_info_destroy, NULL);
     luaT_newmetatable(L, nerv_chunk_data_tname, NULL,
                         NULL, nerv_chunk_data_destroy, NULL);
 }
diff --git a/io/sgd_buffer.lua b/io/sgd_buffer.lua
index dadcf67..bf72744 100644
--- a/io/sgd_buffer.lua
+++ b/io/sgd_buffer.lua
@@ -4,6 +4,10 @@ function SGDBuffer:__init(global_conf, buffer_conf)
     self.gconf = global_conf
     self.buffer_size = math.floor(buffer_conf.buffer_size /
                                 global_conf.batch_size) * global_conf.batch_size
+    self.randomize = buffer_conf.randomize
+    if self.randomize == nil then
+        self.randomize = false
+    end
     self.head = 0
     self.tail = 0
     self.readers = {}
@@ -35,7 +39,9 @@ function SGDBuffer:saturate()
                     nerv.error("buffer size is too small to contain leftovers")
                 end
                 buff.data:copy_from(buff.leftover, 0, lrow)
+                buff.leftover = nil
             end
+            nerv.utils.printf("leftover: %d\n", lrow)
             reader.tail = lrow
             reader.has_leftover = false
         end
@@ -73,6 +79,8 @@ function SGDBuffer:saturate()
         end
         self.tail = math.min(self.tail, reader.tail)
     end
+    self.rand_map = nerv.MMatrixInt.perm_gen(self.tail) -- generate shuffled index
+    collectgarbage("collect")
     return self.tail >= self.gconf.batch_size
 end
 
@@ -90,7 +98,11 @@ function SGDBuffer:get_data()
     for i, reader in ipairs(self.readers) do
         for id, buff in pairs(reader.buffs) do
             local batch = self.gconf.cumat_type(batch_size, buff.width)
-            batch:copy_fromh(buff.data, self.head, self.head + batch_size)
+            if self.randomize then
+                batch:copy_rows_fromh_by_idx(buff.data, self.rand_map, self.head)
+            else
+                batch:copy_fromh(buff.data, self.head, self.head + batch_size)
+            end
             res[id] = batch
         end
     end
diff --git a/layer/softmax_ce.lua b/layer/softmax_ce.lua
index cf98c45..cd57010 100644
--- a/layer/softmax_ce.lua
+++ b/layer/softmax_ce.lua
@@ -17,6 +17,7 @@ function SoftmaxCELayer:init()
         nerv.error("mismatching dimensions of previous network output and labels")
     end
     self.total_ce = 0.0
+    self.total_correct = 0
     self.total_frames = 0
 end
 
@@ -27,7 +28,7 @@ end
 function SoftmaxCELayer:propagate(input, output)
     local soutput = input[1]:create()  -- temporary value for calc softmax
     self.soutput = soutput
-    soutput:softmax(input[1])
+    local classified = soutput:softmax(input[1])
     local ce = soutput:create()
     ce:log_elem(soutput)
     local label = input[2]
@@ -38,6 +39,7 @@ function SoftmaxCELayer:propagate(input, output)
     -- add total ce
     self.total_ce = self.total_ce - ce:rowsum():colsum()[0]
     self.total_frames = self.total_frames + soutput:nrow()
+    self.total_correct = self.total_correct + classified:colsame(input[2])[0]
 end
 
 function SoftmaxCELayer:back_propagate(next_bp_err, bp_err, input, output)
diff --git a/matrix/cuda_helper.h b/matrix/cuda_helper.h
index cedc643..5e5f2ad 100644
--- a/matrix/cuda_helper.h
+++ b/matrix/cuda_helper.h
@@ -1,17 +1,23 @@
 #ifndef NERV_CUDA_HELPER_H
 #define NERV_CUDA_HELPER_H
+#include "cuda.h"
+#include "cuda_runtime.h"
+#include "driver_types.h"
+#include "cublas_v2.h"
 #define CUBLAS_SAFE_CALL(call) \
     do { \
         cublasStatus_t  err = (call); \
         if (err != CUBLAS_STATUS_SUCCESS) \
-            nerv_error(L, "cumatrix cublas error: %s", cublasGetErrorString(err)); \
+            nerv_error(L, "cumatrix cublas error: %s at %s:%d", \
+                        cublasGetErrorString(err), __FILE__, __LINE__); \
     } while (0)
 
 #define CUDA_SAFE_CALL(call) \
     do { \
         cudaError_t err = (call); \
         if (err != cudaSuccess) \
-            nerv_error(L, "cumatrix CUDA error: %s", cudaGetErrorString(err)); \
+            nerv_error(L, "cumatrix CUDA error: %s at %s:%d", \
+                            cudaGetErrorString(err), __FILE__, __LINE__); \
     } while (0)
 
 #define CUDA_SAFE_SYNC_CALL(call) \
@@ -52,4 +58,20 @@ static const char *cublasGetErrorString(cublasStatus_t err) {
     }
     return "<unknown>";
 }
+
+#define PROFILE_START \
+    do { \
+        cudaEvent_t start, stop; \
+        cudaEventCreate(&start); \
+        cudaEventCreate(&stop); \
+        cudaEventRecord(start, 0);
+#define PROFILE_STOP \
+        cudaEventRecord(stop, 0); \
+        cudaEventSynchronize(stop); \
+        float milliseconds = 0; \
+        cudaEventElapsedTime(&milliseconds, start, stop); \
+        accu_profile(__func__, milliseconds / 1000); \
+    } while (0);
+
+#define PROFILE_END 
 #endif
diff --git a/matrix/cukernel.h b/matrix/cukernel.h
index 7d2168e..23398c8 100644
--- a/matrix/cukernel.h
+++ b/matrix/cukernel.h
@@ -5,7 +5,9 @@ void cudak_(cuda_sigmoid)(const Matrix *a, Matrix *b);
 void cudak_(cuda_sigmoid_grad)(const Matrix *output, const Matrix *err, Matrix *nerr);
 void cudak_(cuda_rowsum)(const Matrix *a, Matrix *b);
 void cudak_(cuda_rowmax)(const Matrix *a, Matrix *b);
+void cudak_(cuda_rowmax_idx)(const Matrix *a, Matrix *b, Matrix *idx);
 void cudak_(cuda_colsum)(const Matrix *a, Matrix *b);
+void cudak_(cuda_colsame)(const Matrix *a, const Matrix *ref, Matrix *b);
 void cudak_(cuda_softmax_denominator)(const Matrix *a, const Matrix *max, Matrix *b);
 void cudak_(cuda_softmax_final)(const Matrix *a, const Matrix *max, const Matrix *deno, Matrix *b);
 void cudak_(cuda_add_row)(const Matrix *a, Matrix *b, double beta);
diff --git a/matrix/cumatrix.c b/matrix/cumatrix.c
index 51a3681..4ebc5ff 100644
--- a/matrix/cumatrix.c
+++ b/matrix/cumatrix.c
@@ -1,4 +1,38 @@
 #define NERV_GENERIC_CUMATRIX
+#include "../common.h"
+#include "cuda_helper.h"
+static cublasHandle_t cublas_handle;
+static HashMap *profile;
+
+int print_profile(lua_State *L) {
+    size_t i;
+    fprintf(stderr, "*** [nerv cumatrix profile] **\n");
+    for (i = 0; i < profile->size; i++)
+    {
+        HashNode *ptr;
+        for (ptr = profile->bucket[i]; ptr; ptr = ptr->next)
+        {
+            fprintf(stderr, "%s:\t%.6f\n", ptr->key, *(float *)ptr->val);
+        }
+    }
+    return 0;
+}
+
+int clear_profile(lua_State *L) {
+    hashmap_clear(profile);
+    return 0;
+}
+
+void accu_profile(const char *name, float delta) {
+    float *val = hashmap_getval(profile, name);
+    if (!val)
+    {
+        val = malloc(sizeof(float));
+        *val = 0;
+        hashmap_setval(profile, name, val);
+    }
+    *val += delta;
+}
 
 #define MATRIX_USE_FLOAT
 #define cuda_matrix_(NAME) cuda_matrix_float_##NAME
diff --git a/matrix/generic/cukernel.cu b/matrix/generic/cukernel.cu
index 05a1e78..fdab356 100644
--- a/matrix/generic/cukernel.cu
+++ b/matrix/generic/cukernel.cu
@@ -3,6 +3,7 @@
 #include <stdio.h>
 #include "matrix.h"
 #include "cuda.h"
+#include "float.h"
 #define CUDA_THREADS_N 16
 #define CUDA_THREADS_NN ((CUDA_THREADS_N) * (CUDA_THREADS_N))
 #define CEIL_DIV(a, b) (((a) + (b) - 1) / (b))
@@ -11,9 +12,12 @@ __global__ void cudak_(log_elem)(const MATRIX_ELEM *a, MATRIX_ELEM *b,
     int j = blockIdx.x * blockDim.x + threadIdx.x;
     int i = blockIdx.y * blockDim.y + threadIdx.y;
     long idx;
+    MATRIX_ELEM tmp;
     if (i >= nrow || j >= ncol) return;
     idx = j + i * stride;
-    b[idx] = log(a[idx]);
+    tmp = a[idx];
+    if(tmp < FLT_MIN) tmp = FLT_MIN;
+    b[idx] = log(tmp);
 }
 
 __global__ void cudak_(mul_elem)(const MATRIX_ELEM *a, const MATRIX_ELEM *b,
@@ -61,9 +65,9 @@ __global__ void cudak_(softmax_final)(const MATRIX_ELEM *a, MATRIX_ELEM *b,
 }
 
 __global__ void cudak_(block_reduce_rowsum)(const MATRIX_ELEM *input,
-                                MATRIX_ELEM *output,
-                                const int istride, const int ostride,
-                                const int n) {
+                                            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;
@@ -96,6 +100,26 @@ __global__ void cudak_(block_reduce_colsum)(const MATRIX_ELEM *input,
         output[blockIdx.x + ostride * blockIdx.y] = cudak_(arr)[0];
 }
 
+__global__ void cudak_(block_reduce_colsame)(const MATRIX_ELEM *input,
+                                            const MATRIX_ELEM *ref_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] == \
+                                        ref_input[blockIdx.x + istride * i]) ? 1.0 : 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,
@@ -117,9 +141,9 @@ __global__ void cudak_(block_reduce_softmax_rowsum)(const MATRIX_ELEM *input,
 }
 
 __global__ void cudak_(block_reduce_rowmax)(const MATRIX_ELEM *input,
-                                MATRIX_ELEM *output,
-                                const int istride, const int ostride,
-                                const int n) {
+                                            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;
@@ -129,8 +153,9 @@ __global__ void cudak_(block_reduce_rowmax)(const MATRIX_ELEM *input,
         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;
+                        r = cudak_(arr)[threadIdx.x + offset];
+            if (r > l)
+                cudak_(arr)[threadIdx.x] = r;
         }
         __syncthreads();
     }
@@ -138,6 +163,40 @@ __global__ void cudak_(block_reduce_rowmax)(const MATRIX_ELEM *input,
         output[blockIdx.x + ostride * blockIdx.y] = cudak_(arr)[0];
 }
 
+__global__ void cudak_(block_reduce_rowmax_idx)(const MATRIX_ELEM *input,
+                                                const MATRIX_ELEM *idx_input,
+                                                MATRIX_ELEM *output,
+                                                MATRIX_ELEM *idx_output,
+                                                const int istride, const int ostride,
+                                                const int n) {
+    extern __shared__ MATRIX_ELEM cudak_(arr)[];
+    MATRIX_ELEM *arr_val = cudak_(arr);
+    MATRIX_ELEM *arr_idx = arr_val + blockDim.x;
+    int j = blockIdx.x * blockDim.x + threadIdx.x;
+    arr_val[threadIdx.x] = j < n ? input[j + istride * blockIdx.y] : 0;
+    arr_idx[threadIdx.x] = j < n ? idx_input[j + istride * blockIdx.y] : 0;
+    __syncthreads();
+    for (int offset = blockDim.x >> 1;  offset; offset >>= 1)
+    {
+        if (threadIdx.x < offset)
+        {
+            MATRIX_ELEM l = arr_val[threadIdx.x],
+                        r = arr_val[threadIdx.x + offset];
+            if (r > l)
+            {
+                arr_val[threadIdx.x] = r;
+                arr_idx[threadIdx.x] = arr_idx[threadIdx.x + offset];
+            }
+        }
+        __syncthreads();
+    }
+    if (threadIdx.x == 0)
+    {
+        output[blockIdx.x + ostride * blockIdx.y] = arr_val[0];
+        idx_output[blockIdx.x + ostride * blockIdx.y] = arr_idx[0];
+    }
+}
+
 __global__ void cudak_(add_row)(const MATRIX_ELEM *a, MATRIX_ELEM *b,
                                 int nrow, int ncol, int stride, double beta) {
     int j = blockIdx.x * blockDim.x + threadIdx.x;
@@ -196,6 +255,14 @@ __global__ void cudak_(decompress)(const MATRIX_ELEM *a, MATRIX_ELEM *b,
     b[lrintf(a[j + i * stride_a]) + i * stride_b] = 1.0;
 }
 
+__global__ void cudak_(gen_col_idx)(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;
+    if (i >= nrow || j >= ncol) return;
+    b[j + i * stride] = j;
+}
+
 extern "C" {
 #include "../cukernel.h"
     void cudak_(cuda_log_elem)(const Matrix *a, Matrix *b) {
@@ -261,10 +328,32 @@ extern "C" {
         cudaFree(res);
     }
 
+    void cudak_(cuda_colsame)(const Matrix *a, const Matrix *ref, Matrix *b) {
+        dim3 block(1, CUDA_THREADS_NN);
+        int nrow = a->nrow;
+        int blocks_per_col = CEIL_DIV(nrow, block.y);
+        dim3 grid(a->ncol, blocks_per_col);
+        MATRIX_ELEM *res;
+        size_t stride;
+        cudaMallocPitch(&res, &stride, a->ncol * sizeof(MATRIX_ELEM), blocks_per_col);
+        cudak_(block_reduce_colsame)<<<grid, block, block.y * sizeof(MATRIX_ELEM)>>> \
+            (MATRIX_ELEM_PTR(a), MATRIX_ELEM_PTR(ref), res,
+             a->stride / sizeof(MATRIX_ELEM), stride / sizeof(MATRIX_ELEM),
+             nrow);
+        nrow = blocks_per_col;
+        assert((unsigned long)nrow <= block.y);
+        grid.y = 1;
+        cudak_(block_reduce_colsum)<<<grid, block, block.y * sizeof(MATRIX_ELEM)>>> \
+            (res, MATRIX_ELEM_PTR(b),
+             stride / sizeof(MATRIX_ELEM), b->stride / sizeof(MATRIX_ELEM),
+             nrow);
+        cudaFree(res);
+    }
+
     void cudak_(cuda_colsum)(const Matrix *a, Matrix *b) {
         dim3 block(1, CUDA_THREADS_NN);
         int nrow = a->nrow;
-        int blocks_per_col = CEIL_DIV(nrow, block.x);
+        int blocks_per_col = CEIL_DIV(nrow, block.y);
         dim3 grid(a->ncol, blocks_per_col);
         MATRIX_ELEM *res;
         size_t stride;
@@ -344,6 +433,35 @@ extern "C" {
         cudaFree(res);
     }
 
+    void cudak_(cuda_rowmax_idx)(const Matrix *a, Matrix *b, Matrix *b_idx) {
+        dim3 block(CUDA_THREADS_NN, 1);
+        int ncol = a->ncol;
+        int blocks_per_row = CEIL_DIV(ncol, block.x);
+        dim3 grid(blocks_per_row, a->nrow);
+        MATRIX_ELEM *a_idx, *res, *res_idx;
+        size_t stride;
+        cudaMallocPitch(&a_idx, &stride, a->stride, a->nrow);
+        cudak_(gen_col_idx)<<<grid, block>>>(a_idx, a->nrow, ncol, stride / sizeof(MATRIX_ELEM));
+        cudaMallocPitch(&res, &stride, blocks_per_row * sizeof(MATRIX_ELEM), a->nrow);
+        cudaMallocPitch(&res_idx, &stride, blocks_per_row * sizeof(MATRIX_ELEM), a->nrow);
+        cudak_(block_reduce_rowmax_idx)<<<grid, block,
+                                        2 * block.x * sizeof(MATRIX_ELEM)>>> \
+            (MATRIX_ELEM_PTR(a), a_idx, res, res_idx,
+             a->stride / sizeof(MATRIX_ELEM), stride / sizeof(MATRIX_ELEM),
+             ncol);
+        cudaFree(a_idx);
+        ncol = blocks_per_row;
+        assert((unsigned long)ncol <= block.x);
+        grid.x = 1;
+        cudak_(block_reduce_rowmax_idx)<<<grid, block,
+                                        2 * block.x * sizeof(MATRIX_ELEM)>>> \
+            (res, res_idx, MATRIX_ELEM_PTR(b), MATRIX_ELEM_PTR(b_idx),
+             stride / sizeof(MATRIX_ELEM), b->stride / sizeof(MATRIX_ELEM),
+             ncol);
+        cudaFree(res);
+        cudaFree(res_idx);
+    }
+
     /* in-place calc */
     void cudak_(cuda_add_row)(const Matrix *a, Matrix *b, double beta) {
         dim3 threadsPerBlock(CUDA_THREADS_N, CUDA_THREADS_N);
diff --git a/matrix/generic/cumatrix.c b/matrix/generic/cumatrix.c
index 373fc42..8e7d34f 100644
--- a/matrix/generic/cumatrix.c
+++ b/matrix/generic/cumatrix.c
@@ -11,15 +11,11 @@
 #define MATRIX_BASE_TNAME nerv_matrix_cuda_tname
 #define NERV_GENERIC_MATRIX
 #define NERV_GENERIC_CUKERNEL
+#define PROFILE_HASHMAP_SIZE 123457
 #include "../../common.h"
 #include "../cukernel.h"
-#include "cuda.h"
-#include "cuda_runtime.h"
-#include "driver_types.h"
-#include "cublas_v2.h"
 #include "../cuda_helper.h"
-
-static cublasHandle_t cublas_handle;
+#include <string.h>
 
 Matrix *nerv_matrix_(new_)(lua_State *L, long nrow, long ncol);
 void nerv_matrix_(data_free)(lua_State *L, Matrix *self);
@@ -27,6 +23,7 @@ void nerv_matrix_(data_free)(lua_State *L, Matrix *self);
 static void nerv_matrix_(add_)(lua_State *L, const Matrix *a, const Matrix *b,
                                 const Matrix *c,
                                 MATRIX_ELEM alpha, MATRIX_ELEM beta) {
+    PROFILE_START
     CUBLAS_SAFE_CALL(
             NERV_CUBLAS_(geam)(cublas_handle, CUBLAS_OP_N, CUBLAS_OP_N,
                 a->ncol, a->nrow,
@@ -35,6 +32,7 @@ static void nerv_matrix_(add_)(lua_State *L, const Matrix *a, const Matrix *b,
                 &beta,
                 MATRIX_ELEM_PTR(b), b->stride / sizeof(MATRIX_ELEM),
                 MATRIX_ELEM_PTR(c), c->stride / sizeof(MATRIX_ELEM)));
+    PROFILE_STOP
 }
 
 static int nerv_matrix_(add)(lua_State *L) {
@@ -75,6 +73,7 @@ static int nerv_matrix_(mul)(lua_State *L) {
         nerv_error(L, "Wrong dimension of multipliers");
 /*    MATRIX_ELEM alpha = 1.0f, beta = 0.0f; */
     /* Because matrix in Nerv is row-major, here b comes first */
+    PROFILE_START
     CUBLAS_SAFE_CALL(
             NERV_CUBLAS_(gemm)(cublas_handle, tb, ta,
                 bn, am, bm,
@@ -83,6 +82,7 @@ static int nerv_matrix_(mul)(lua_State *L) {
                 MATRIX_ELEM_PTR(a), a->stride / sizeof(MATRIX_ELEM),
                 &beta,
                 MATRIX_ELEM_PTR(c), c->stride / sizeof(MATRIX_ELEM)));
+    PROFILE_STOP
     return 0;
 }
 
@@ -97,7 +97,9 @@ static int nerv_matrix_(sigmoid)(lua_State *L) {
     Matrix *a = luaT_checkudata(L, 1, nerv_matrix_(tname));
     Matrix *b = luaT_checkudata(L, 2, nerv_matrix_(tname));
     CHECK_SAME_DIMENSION(a, b);
+    PROFILE_START
     cudak_(cuda_sigmoid)(b, a);
+    PROFILE_STOP
     return 0;
 }
 
@@ -107,30 +109,38 @@ static int nerv_matrix_(sigmoid_grad)(lua_State *L) {
     Matrix *output = luaT_checkudata(L, 3, nerv_matrix_(tname));
     CHECK_SAME_DIMENSION(nerr, err);
     CHECK_SAME_DIMENSION(nerr, output);
+    PROFILE_START
     cudak_(cuda_sigmoid_grad)(output, err, nerr);
+    PROFILE_STOP
     return 0;
 }
 
 static int nerv_matrix_(softmax)(lua_State *L) {
     Matrix *a = luaT_checkudata(L, 2, nerv_matrix_(tname));
     Matrix *b = luaT_checkudata(L, 1, nerv_matrix_(tname));
-    Matrix *max;
+    Matrix *max, *max_idx;
     Matrix *dno;
     CHECK_SAME_DIMENSION(a, b);
     max = nerv_matrix_(new_)(L, a->nrow, 1);
+    max_idx = nerv_matrix_(new_)(L, a->nrow, 1);
     dno = nerv_matrix_(new_)(L, a->nrow, 1);
-    cudak_(cuda_rowmax)(a, max);
+    PROFILE_START
+    cudak_(cuda_rowmax_idx)(a, max, max_idx);
     cudak_(cuda_softmax_denominator)(a, max, dno);
     cudak_(cuda_softmax_final)(a, max, dno, b);
+    PROFILE_STOP
     nerv_matrix_(data_free)(L, max);
     nerv_matrix_(data_free)(L, dno);
-    return 0;
+    luaT_pushudata(L, max_idx, nerv_matrix_(tname));
+    return 1;
 }
 
 static int nerv_matrix_(rowsum)(lua_State *L) {
     Matrix *a = luaT_checkudata(L, 1, nerv_matrix_(tname));
     Matrix *b = nerv_matrix_(new_)(L, a->nrow, 1);
+    PROFILE_START
     cudak_(cuda_rowsum)(a, b);
+    PROFILE_STOP
     luaT_pushudata(L, b, nerv_matrix_(tname));
     return 1;
 }
@@ -138,7 +148,21 @@ static int nerv_matrix_(rowsum)(lua_State *L) {
 static int nerv_matrix_(colsum)(lua_State *L) {
     Matrix *a = luaT_checkudata(L, 1, nerv_matrix_(tname));
     Matrix *b = nerv_matrix_(new_)(L, 1, a->ncol);
+    PROFILE_START
     cudak_(cuda_colsum)(a, b);
+    PROFILE_STOP
+    luaT_pushudata(L, b, nerv_matrix_(tname));
+    return 1;
+}
+
+static int nerv_matrix_(colsame)(lua_State *L) {
+    Matrix *a = luaT_checkudata(L, 1, nerv_matrix_(tname));
+    Matrix *ref = luaT_checkudata(L, 2, nerv_matrix_(tname));
+    Matrix *b = nerv_matrix_(new_)(L, 1, a->ncol);
+    CHECK_SAME_DIMENSION(a, ref);
+    PROFILE_START
+    cudak_(cuda_colsame)(a, ref, b);
+    PROFILE_STOP
     luaT_pushudata(L, b, nerv_matrix_(tname));
     return 1;
 }
@@ -146,11 +170,24 @@ static int nerv_matrix_(colsum)(lua_State *L) {
 static int nerv_matrix_(rowmax)(lua_State *L) {
     Matrix *a = luaT_checkudata(L, 1, nerv_matrix_(tname));
     Matrix *b = nerv_matrix_(new_)(L, a->nrow, 1);
+    PROFILE_START
     cudak_(cuda_rowmax)(a, b);
+    PROFILE_STOP
     luaT_pushudata(L, b, nerv_matrix_(tname));
     return 1;
 }
 
+static int nerv_matrix_(rowmax_idx)(lua_State *L) {
+    Matrix *a = luaT_checkudata(L, 1, nerv_matrix_(tname));
+    Matrix *b = nerv_matrix_(new_)(L, a->nrow, 1);
+    Matrix *idx = nerv_matrix_(new_)(L, a->nrow, 1);
+    PROFILE_START
+    cudak_(cuda_rowmax_idx)(a, b, idx);
+    PROFILE_STOP
+    luaT_pushudata(L, b, nerv_matrix_(tname));
+    luaT_pushudata(L, idx, nerv_matrix_(tname));
+    return 2;
+}
 
 static int nerv_matrix_(add_row)(lua_State *L) {
     Matrix *a = luaT_checkudata(L, 2, nerv_matrix_(tname));
@@ -160,14 +197,18 @@ static int nerv_matrix_(add_row)(lua_State *L) {
         nerv_error(L, "the number of columns is not the same");
     if (a->nrow != 1)
         nerv_error(L, "a row vector is expected");
+    PROFILE_START
     cudak_(cuda_add_row)(a, b, beta);
+    PROFILE_STOP
     return 0;
 }
 
 static int nerv_matrix_(fill)(lua_State *L) {
     Matrix *self = luaT_checkudata(L, 1, nerv_matrix_(tname));
     double val = luaL_checknumber(L, 2);
+    PROFILE_START
     cudak_(cuda_fill)(self, val);
+    PROFILE_STOP
     return 0;
 }
 
@@ -183,11 +224,13 @@ static int nerv_matrix_(copy_fromd)(lua_State *L) {
         nerv_error(L, "invalid copy interval");
     if (a->ncol != b->ncol)
         nerv_error(L, "matrices should be of the same dimension");
+    PROFILE_START
     CUDA_SAFE_SYNC_CALL(
             cudaMemcpy2D(MATRIX_ROW_PTR(a, a_begin), a->stride,
                 MATRIX_ROW_PTR(b, b_begin), b->stride,
                 sizeof(MATRIX_ELEM) * b->ncol, b_end - b_begin,
                 cudaMemcpyDeviceToDevice));
+    PROFILE_STOP
     return 0;
 }
 
@@ -204,11 +247,13 @@ static int nerv_matrix_(copy_fromh)(lua_State *L) {
         nerv_error(L, "invalid copy interval");
     if (a->ncol != b->ncol)
         nerv_error(L, "matrices should be of the same dimension");
+    PROFILE_START
     CUDA_SAFE_SYNC_CALL(
             cudaMemcpy2D(MATRIX_ROW_PTR(a, a_begin), a->stride,
                 MATRIX_ROW_PTR(b, b_begin), b->stride,
                 sizeof(MATRIX_ELEM) * b->ncol, b_end - b_begin,
                 cudaMemcpyHostToDevice));
+    PROFILE_STOP
     return 0;
 }
 
@@ -224,11 +269,13 @@ static int nerv_matrix_(copy_toh)(lua_State *L) {
         nerv_error(L, "invalid copy interval");
     if (b->ncol != a->ncol)
         nerv_error(L, "matrices should be of the same dimension");
+    PROFILE_START
     CUDA_SAFE_SYNC_CALL(
             cudaMemcpy2D(MATRIX_ROW_PTR(b, b_begin), b->stride,
                 MATRIX_ROW_PTR(a, a_begin), a->stride,
                 sizeof(MATRIX_ELEM) * a->ncol, a_end - a_begin,
                 cudaMemcpyDeviceToHost));
+    PROFILE_STOP
     return 0;
 }
 
@@ -237,6 +284,7 @@ static int nerv_matrix_(trans)(lua_State *L) {
     Matrix *b = nerv_matrix_(new_)(L, a->ncol, a->nrow);
     MATRIX_ELEM alpha = 1, beta = 0;
     /* FIXME: possible memory leak when lua error is raised */
+    PROFILE_START
     CUBLAS_SAFE_CALL(
             NERV_CUBLAS_(geam)(cublas_handle, CUBLAS_OP_T, CUBLAS_OP_T,
                 a->nrow, a->ncol,
@@ -245,6 +293,7 @@ static int nerv_matrix_(trans)(lua_State *L) {
                 &beta,
                 MATRIX_ELEM_PTR(a), a->stride / sizeof(MATRIX_ELEM),
                 MATRIX_ELEM_PTR(b), b->stride / sizeof(MATRIX_ELEM)));
+    PROFILE_STOP
     luaT_pushudata(L, b, nerv_matrix_(tname));
     return 1;
 }
@@ -255,7 +304,9 @@ static int nerv_matrix_(mul_elem)(lua_State *L) {
     Matrix *c = luaT_checkudata(L, 1, nerv_matrix_(tname));
     CHECK_SAME_DIMENSION(a, b);
     CHECK_SAME_DIMENSION(a, c);
+    PROFILE_START
     cudak_(cuda_mul_elem)(a, b, c);
+    PROFILE_STOP
     return 0;
 }
 
@@ -263,7 +314,9 @@ 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);
+    PROFILE_START
     cudak_(cuda_log_elem)(a, b);
+    PROFILE_STOP
     return 0;
 }
 
@@ -274,8 +327,10 @@ static int nerv_matrix_(decompress)(lua_State *L) {
     if (a->ncol != 1)
         nerv_error(L, "the compressed matrix must be a column vector");
     b = nerv_matrix_(new_)(L, a->nrow, orig_col);
+    PROFILE_START
     cudak_(cuda_fill)(b, 0.0);
     cudak_(cuda_decompress)(a, b);
+    PROFILE_STOP
     luaT_pushudata(L, b, nerv_matrix_(tname));
     return 1;
 }
@@ -285,21 +340,25 @@ static int nerv_matrix_(copy_rows_fromh_by_idx)(lua_State *L) {
     Matrix *a = luaT_checkudata(L, 1, nerv_matrix_(tname));
     Matrix *b = luaT_checkudata(L, 2, MATRIX_CUMATRIX_HOST_TNAME);
     Matrix *idx = luaT_checkudata