fastnn compile

8 files changed, 624 insertions, 43 deletions

diff --git a/fastnn/example/asgd_sds_trainer.lua b/fastnn/example/asgd_sds_trainer.lua
index 44611b2..cf1c7a6 100644
--- a/fastnn/example/asgd_sds_trainer.lua
+++ b/fastnn/example/asgd_sds_trainer.lua
@@ -1,19 +1,22 @@
-package.path="/home/slhome/wd007/.luarocks/share/lua/5.1/?.lua;/home/slhome/wd007/.luarocks/share/lua/5.1/?/init.lua;/sgfs/users/wd007/src/nerv/install/share/lua/5.1/?.lua;/sgfs/users/wd007/src/nerv/install/share/lua/5.1/?/init.lua;"..package.path; 
-package.cpath="/home/slhome/wd007/.luarocks/lib/lua/5.1/?.so;/sgfs/users/wd007/src/nerv/install/lib/lua/5.1/?.so;"..package.cpath
+
+NERV_ROOT = "/sgfs/users/wd007/src/nerv-2"
+
+env = string.format([[
+package.path="/home/slhome/wd007/.luarocks/share/lua/5.1/?.lua;/home/slhome/wd007/.luarocks/share/lua/5.1/?/init.lua;%s/install/share/lua/5.1/?.lua;%s/install/share/lua/5.1/?/init.lua;"..package.path; 
+package.cpath="/home/slhome/wd007/.luarocks/lib/lua/5.1/?.so;%s/install/lib/lua/5.1/?.so;"..package.cpath
 local k,l,_=pcall(require,"luarocks.loader") _=k and l.add_context("nerv","scm-1")
+]], NERV_ROOT, NERV_ROOT, NERV_ROOT)
+
+loadstring(env)()
+
 require 'nerv'
 
 require 'fastnn'
 require 'libhtkio'
 require 'threads'
 
-dofile("fastnn/fastnn_baseline.lua")
+dofile("fastnn/example/fastnn_baseline.lua")
 
-env = string.format([[
-package.path="/home/slhome/wd007/.luarocks/share/lua/5.1/?.lua;/home/slhome/wd007/.luarocks/share/lua/5.1/?/init.lua;/sgfs/users/wd007/src/nerv/install/share/lua/5.1/?.lua;/sgfs/users/wd007/src/nerv/install/share/lua/5.1/?/init.lua;"..package.path; 
-package.cpath="/home/slhome/wd007/.luarocks/lib/lua/5.1/?.so;/sgfs/users/wd007/src/nerv/install/lib/lua/5.1/?.so;"..package.cpath
-local k,l,_=pcall(require,"luarocks.loader") _=k and l.add_context("nerv","scm-1")
-]])
 
 
 train_thread_code = [[
@@ -21,7 +24,9 @@ train_thread_code = [[
 
 require 'nerv'
 require 'fastnn'
-dofile("fastnn/fastnn_baseline.lua")
+require 'libhtkio'
+
+dofile("fastnn/example/fastnn_baseline.lua")
 os.execute("export MALLOC_CHECK_=0")
 
 local thread_idx = %d
@@ -49,39 +54,35 @@ else
 	gconf.tr_scp = scp_file
 end
 
+share_mutex:lock()
+
 gconf.randomize = bp
 gconf.lrate = lrate
 gconf.batch_size = batch_size
-gconf.network[1] = nnet_in
-nerv.info_stderr("input network: %%s", gconf.network[1])
+gconf.initialized_param[2] = nnet_in
+nerv.info_stderr("input network: %%s", gconf.initialized_param[2])
 --nerv.info_stderr(gconf.randomize)
 nerv.info_stderr("input batch_size: %%d", gconf.batch_size)
 nerv.info_stderr("input scp_file: %%s", scp_file)
 nerv.info_stderr("input lrate: %%f", gconf.lrate)
 
 
-share_mutex:lock()
+
 share_gpu:select_gpu()
 
 nerv.context = nerv.CCuContext()
 --print(nerv.context)
 
-nerv.info_stderr("thread %%d loading transf ...", thread_idx)
-local param_transf_repo = nerv.ParamRepo()
-param_transf_repo:import(gconf.transf, nil, gconf)
-local transf_node_repo = make_transf_node_repo(param_transf_repo)
-local transf_layer_repo = make_transf_link_repo(transf_node_repo, param_transf_repo)
-local transf = transf_layer_repo:get_layer("global_transf")
-
-nerv.info_stderr("thread %%d loading network ...", thread_idx)
-local param_network_repo = nerv.ParamRepo()
-param_network_repo:import(gconf.network, nil, gconf)
-local network_node_repo = make_network_node_repo(param_network_repo)
-local network_layer_repo = make_network_link_repo(network_node_repo, param_network_repo)
-local network = get_network(network_layer_repo)
+nerv.info_stderr("thread %%d loading parameters ...", thread_idx)
+local param_repo = nerv.ParamRepo()
+param_repo:import(gconf.initialized_param, nil, gconf)
+local layer_repo = make_layer_repo(param_repo)
+local network = get_network(layer_repo)
+local global_transf = get_global_transf(layer_repo)
+
 share_mutex:unlock()
 
-local buffer = make_buffer(make_readers(nil, transf_layer_repo, feat_repo_shareid, data_mutex_shareid))
+local buffer = make_buffer(make_readers(nil, layer_repo, feat_repo_shareid, data_mutex_shareid))
 
 local input_order = get_input_order()
 
@@ -98,21 +99,35 @@ local input_order = get_input_order()
 
 		gconf.cnt = gconf.cnt + 1
 		if gconf.cnt == 2000 then
-			print_stat(network_node_repo)
+			print_stat(layer_repo)
 			gconf.cnt = 0
 		end
 
 		local input = {}
+		
+		for i, e in ipairs(input_order) do
+                	local id = e.id
+                	if data[id] == nil then
+                    		nerv.error("input data %%s not found", id)
+                	end
+                	local transformed
+                	if e.global_transf then
+                    	transformed = nerv.speech_utils.global_transf(data[id],
+                                        global_transf,
+                                        gconf.frm_ext or 0, 0,
+                                        gconf)
+                	else
+                    		transformed = data[id]
+                	end
+	                table.insert(input, transformed)
+            	end
+
+            	local output = {nerv.CuMatrixFloat(gconf.batch_size, 1)}
+            	err_output = {}
+            	for i = 1, #input do
+                	table.insert(err_output, input[i]:create())
+		end
 
-                for i, id in ipairs(input_order) do
-                       if data[id] == nil then
-                                nerv.error("input data %%s not found", id) 
-                        end 
-                        table.insert(input, data[id])
-                end 
-
-		local output = {nerv.CuMatrixFloat(gconf.batch_size, 1)} 
-		err_output = {input[1]:create()}
             	network:propagate(input, output)
 
 		if bp then
@@ -132,7 +147,7 @@ local input_order = get_input_order()
         end 
 
 	--print_stat(network_node_repo)
-	local ce_crit = network_node_repo:get_layer("ce_crit")
+	local ce_crit = layer_repo:get_layer("ce_crit")
 	local xent = fastnn.CXent(ce_crit.total_frames, ce_crit.total_correct, ce_crit.total_ce, ce_crit.total_ce)
 
 	share_master:LockModel()
@@ -236,8 +251,8 @@ min_iter = 1
 max_iter = 20
 min_halving = 0
 gconf.batch_size = 256
-pf0 = get_filename(gconf.network[1])
-nnet_in = gconf.network[1]
+pf0 = get_filename(gconf.initialized_param[2])
+nnet_in = gconf.initialized_param[2]
 nnet_out = ""
 sds_scp = "tr_sds_"..string.format("%.4d", math.random()*10000)..".scp"  --"tr_sds.scp"
 sds_factor = 0.4
diff --git a/fastnn/example/fastnn_baseline.lua b/fastnn/example/fastnn_baseline.lua
new file mode 100644
index 0000000..6e774de
--- /dev/null
+++ b/fastnn/example/fastnn_baseline.lua
@@ -0,0 +1,258 @@
+require 'htk_io'
+
+gconf = {lrate = 0.2, wcost = 1e-6, momentum = 0.9,
+        cumat_type = nerv.CuMatrixFloat,
+        mmat_type = nerv.MMatrixFloat,
+        frm_ext = 5,
+        frm_trim = 5,
+	batch_size = 256,
+	buffer_size = 81920,
+	rearrange = true,
+        tr_scp = "/sgfs/users/wd007/asr/baseline_chn_50h/finetune/finetune_baseline/train.scp",
+        cv_scp = "/sgfs/users/wd007/asr/baseline_chn_50h/finetune/finetune_baseline/train_cv.scp",
+        htk_conf = "/sgfs/users/wd007/asr/baseline_chn_50h/finetune/finetune_baseline/fbank_d_a_z.conf",
+	initialized_param = {"/sgfs/users/wd007/src/nerv/tools/nerv.global.transf",
+                             "/sgfs/users/wd007/src/nerv/tools/nerv.svd0.55_3000h_iter1.init"},
+        debug = false}
+
+function make_layer_repo(param_repo) 
+    local layer_repo = nerv.LayerRepo( 
+    {
+        -- global transf
+        ["nerv.BiasLayer"] =
+        {   
+            blayer1 = {{bias = "bias1"}, {dim_in = {1320}, dim_out = {1320}}},
+        },  
+        ["nerv.WindowLayer"] =
+        {   
+            wlayer1 = {{window = "window1"}, {dim_in = {1320}, dim_out = {1320}}},
+	},
+        -- biased linearity
+        ["nerv.AffineLayer"] =
+        {
+            affine0 = {{ltp = "affine0_ltp", bp = "affine0_bp"},
+            {dim_in = {1320}, dim_out = {2048}}},
+            affine1 = {{ltp = "affine1_ltp", bp = "affine1_bp"},
+            {dim_in = {2048}, dim_out = {367}}},
+            affine2 = {{ltp = "affine2_ltp", bp = "affine2_bp"},
+            {dim_in = {367}, dim_out = {2048}}},
+            affine3 = {{ltp = "affine3_ltp", bp = "affine3_bp"},
+            {dim_in = {2048}, dim_out = {408}}},
+            affine4 = {{ltp = "affine4_ltp", bp = "affine4_bp"},
+            {dim_in = {408}, dim_out = {2048}}},
+            affine5 = {{ltp = "affine5_ltp", bp = "affine5_bp"},
+            {dim_in = {2048}, dim_out = {368}}},
+            affine6 = {{ltp = "affine6_ltp", bp = "affine6_bp"},
+            {dim_in = {368}, dim_out = {2048}}},
+            affine7 = {{ltp = "affine7_ltp", bp = "affine7_bp"},
+            {dim_in = {2048}, dim_out = {303}}},
+            affine8 = {{ltp = "affine8_ltp", bp = "affine8_bp"},
+            {dim_in = {303}, dim_out = {2048}}},
+            affine9 = {{ltp = "affine9_ltp", bp = "affine9_bp"},
+            {dim_in = {2048}, dim_out = {277}}},
+            affine10 = {{ltp = "affine10_ltp", bp = "affine10_bp"},
+            {dim_in = {277}, dim_out = {2048}}},
+            affine11 = {{ltp = "affine11_ltp", bp = "affine11_bp"},
+            {dim_in = {2048}, dim_out = {361}}},
+            affine12 = {{ltp = "affine12_ltp", bp = "affine12_bp"},
+            {dim_in = {361}, dim_out = {2048}}},
+            affine13 = {{ltp = "affine13_ltp", bp = "affine13_bp"},
+            {dim_in = {2048}, dim_out = {441}}},
+            affine14 = {{ltp = "affine14_ltp", bp = "affine14_bp"},
+            {dim_in = {441}, dim_out = {10092}}},
+	},
+        ["nerv.SigmoidLayer"] =
+        {
+            sigmoid0 = {{}, {dim_in = {2048}, dim_out = {2048}}},
+            sigmoid1 = {{}, {dim_in = {2048}, dim_out = {2048}}},
+            sigmoid2 = {{}, {dim_in = {2048}, dim_out = {2048}}},
+            sigmoid3 = {{}, {dim_in = {2048}, dim_out = {2048}}},
+            sigmoid4 = {{}, {dim_in = {2048}, dim_out = {2048}}},
+            sigmoid5 = {{}, {dim_in = {2048}, dim_out = {2048}}},
+            sigmoid6 = {{}, {dim_in = {2048}, dim_out = {2048}}},
+        },
+        ["nerv.SoftmaxCELayer"] = -- softmax + ce criterion layer for finetune output
+        {
+            ce_crit = {{}, {dim_in = {10092, 1}, dim_out = {1}, compressed = true}}
+        },
+	["nerv.SoftmaxLayer"] = -- softmax for decode output
+        {
+            softmax = {{}, {dim_in = {10092}, dim_out = {10092}}}
+        }
+    }, param_repo, gconf)
+    
+    layer_repo:add_layers(
+    {
+        ["nerv.DAGLayer"] =
+        {
+            global_transf = {{}, {
+                dim_in = {1320}, dim_out = {1320},
+                sub_layers = layer_repo,
+                connections = 
+		{
+                    ["<input>[1]"] = "blayer1[1]",
+                    ["blayer1[1]"] = "wlayer1[1]",
+                    ["wlayer1[1]"] = "<output>[1]"
+                }
+            }},
+	    main = {{}, {
+                dim_in = {1320}, dim_out = {10092},
+                sub_layers = layer_repo,
+                connections = {
+                    ["<input>[1]"] = "affine0[1]",
+                    ["affine0[1]"] = "sigmoid0[1]",
+                    ["sigmoid0[1]"] = "affine1[1]",
+                    ["affine1[1]"] = "affine2[1]",
+                    ["affine2[1]"] = "sigmoid1[1]",
+                    ["sigmoid1[1]"] = "affine3[1]",
+                    ["affine3[1]"] = "affine4[1]",
+                    ["affine4[1]"] = "sigmoid2[1]",
+                    ["sigmoid2[1]"] = "affine5[1]",
+                    ["affine5[1]"] = "affine6[1]",
+                    ["affine6[1]"] = "sigmoid3[1]",
+                    ["sigmoid3[1]"] = "affine7[1]",
+                    ["affine7[1]"] = "affine8[1]",
+                    ["affine8[1]"] = "sigmoid4[1]",
+                    ["sigmoid4[1]"] = "affine9[1]",
+                    ["affine9[1]"] = "affine10[1]",
+                    ["affine10[1]"] = "sigmoid5[1]",
+                    ["sigmoid5[1]"] = "affine11[1]",
+                    ["affine11[1]"] = "affine12[1]",
+                    ["affine12[1]"] = "sigmoid6[1]",
+                    ["sigmoid6[1]"] = "affine13[1]",
+                    ["affine13[1]"] = "affine14[1]",
+                    ["affine14[1]"] = "<output>[1]",
+                }
+            }}
+        }
+    }, param_repo, gconf)
+
+    layer_repo:add_layers(
+    {
+        ["nerv.DAGLayer"] =
+        {
+            ce_output = {{}, {
+                dim_in = {1320, 1}, dim_out = {1},
+                sub_layers = layer_repo,
+                connections = {
+                    ["<input>[1]"] = "main[1]",
+                    ["main[1]"] = "ce_crit[1]",
+                    ["<input>[2]"] = "ce_crit[2]",
+                    ["ce_crit[1]"] = "<output>[1]"
+                }
+            }},
+            softmax_output = {{}, {
+                dim_in = {1320}, dim_out = {10092},
+                sub_layers = layer_repo,
+                connections = {
+                    ["<input>[1]"] = "main[1]",
+                    ["main[1]"] = "softmax[1]",
+                    ["softmax[1]"] = "<output>[1]"
+                }
+            }}
+        }
+    }, param_repo, gconf)
+
+    return layer_repo
+end
+
+
+function get_network(layer_repo)
+    return layer_repo:get_layer("ce_output")
+end
+
+function get_decode_network(layer_repo)
+    return layer_repo:get_layer("softmax_output")
+end
+
+function get_global_transf(layer_repo)
+    return layer_repo:get_layer("global_transf")
+end
+
+
+
+function make_readers(scp_file, layer_repo, feat_repo_shareid, data_mutex_shareid)
+    return {
+                {reader = nerv.TNetReader(gconf,
+                    {   
+                        id = "main_scp",
+                        scp_file = scp_file,
+                        conf_file = gconf.htk_conf,
+                        frm_ext = gconf.frm_ext,
+                        mlfs = { 
+                            phone_state = { 
+                                file = "/sgfs/users/wd007/asr/baseline_chn_50h/finetune/finetune_baseline/ref.mlf",
+                                format = "map",
+                                format_arg = "/sgfs/users/wd007/asr/baseline_chn_50h/finetune/finetune_baseline/dict",
+                                dir = "*/",
+                                ext = "lab"
+                            }   
+                        },  
+                        global_transf = layer_repo:get_layer("global_transf")
+                    }, feat_repo_shareid, data_mutex_shareid), 
+                data = {main_scp = 1320, phone_state = 1}} 
+            }   
+end
+
+function get_feat_id()
+        return {main_scp = true}
+end
+
+
+function make_buffer(readers)
+    return nerv.SGDBuffer(gconf,
+        {
+            buffer_size = gconf.buffer_size,
+            randomize = gconf.randomize,
+            readers = readers,
+            use_gpu = true
+        })
+end
+
+function get_input_order()
+    return {{id = "main_scp", global_transf = true},
+            {id = "phone_state"}}
+end
+
+function get_accuracy(layer_repo)
+    local ce_crit = layer_repo:get_layer("ce_crit")
+    return ce_crit.total_correct / ce_crit.total_frames * 100
+end
+
+function print_stat(layer_repo)
+    local ce_crit = layer_repo:get_layer("ce_crit")
+    nerv.info("*** training stat begin ***")
+    nerv.printf("cross entropy:\t\t%.8f\n", ce_crit.total_ce)
+    nerv.printf("correct:\t\t%d\n", ce_crit.total_correct)
+    nerv.printf("frames:\t\t\t%d\n", ce_crit.total_frames)
+    nerv.printf("err/frm:\t\t%.8f\n", ce_crit.total_ce / ce_crit.total_frames)
+    nerv.printf("accuracy:\t\t%.3f%%\n", get_accuracy(layer_repo))
+    nerv.info("*** training stat end ***")
+end
+
+function print_xent(xent)
+	local totalframes = xent:totalframes()
+	local loss = xent:loss()
+	local correct = xent:correct()
+	nerv.info_stderr("*** training statistics info begin ***")
+	nerv.info_stderr("total frames:\t\t%d", totalframes)
+	nerv.info_stderr("cross entropy:\t%.8f", loss/totalframes)
+	nerv.info_stderr("frame accuracy:\t%.3f%%", 100*correct/totalframes)
+	nerv.info_stderr("*** training statistics info end ***")
+end
+
+function frame_acc(xent)
+	local correct = xent:correct()
+	local totalframes = xent:totalframes()
+	return string.format("%.3f", 100*correct/totalframes)
+end
+
+function print_gconf()
+	nerv.info_stderr("%s \t:= %s", "network", gconf.initialized_param[1])
+	nerv.info_stderr("%s \t:= %s", "transf", gconf.initialized_param[2])
+	nerv.info_stderr("%s \t:= %s", "batch_size", gconf.batch_size)
+	nerv.info_stderr("%s \t:= %s", "buffer_size", gconf.buffer_size)
+	nerv.info_stderr("%s \t:= %s", "lrate", gconf.lrate)
+	nerv.info_stderr("%s \t:= %s", "tr_scp", gconf.tr_scp)
+	nerv.info_stderr("%s \t:= %s", "cv_scp", gconf.cv_scp)
+end
diff --git a/fastnn/lib/ModelSync.c b/fastnn/lib/ModelSync.c
new file mode 100644
index 0000000..bd511ea
--- /dev/null
+++ b/fastnn/lib/ModelSync.c
@@ -0,0 +1,305 @@
+
+#include "ModelSync.h"
+#include "../../nerv/lib/matrix/cuda_helper.h"
+#include "../../nerv/lib/matrix/generic/elem_type.h"
+#include "common.h"
+#include <string.h>
+
+
+ModelSync* ModelSync_new(void)
+{
+	ModelSync *self = (ModelSync*)malloc(sizeof(ModelSync));
+	if (NULL != self)
+	{
+		self->model_mutex = THMutex_new();
+		self->state_mutex = THMutex_new();
+		self->initialized_ = false;
+		self->dim_ = 0;
+		self->pos_ = 0;
+		self->data_ = NULL;
+		self->free_data_ = NULL;
+		self->data_ = NULL;
+		self->refcount = 1;
+		self->threadcount = 0;
+	}
+	return self;
+}
+
+ModelSync* ModelSync_newWithId(long id)
+{
+	ModelSync *self = (ModelSync*)id;
+	__sync_fetch_and_add(&self->refcount, 1);
+	return self;
+}
+
+long ModelSync_id(ModelSync *self) 
+{
+	return (long)(self);
+}
+
+int  ModelSync_lockmodel(ModelSync *self)
+{
+	if(THMutex_lock(self->model_mutex))
+		return 1;
+	return 0;
+}
+	
+int  ModelSync_unlockmodel(ModelSync *self)
+{
+	if(THMutex_unlock(self->model_mutex))
+		return 1;
+        return 0;
+
+}
+int  ModelSync_lockstate(ModelSync *self)
+{
+	if(THMutex_lock(self->state_mutex))
+		return 1;
+        return 0;
+}
+
+int  ModelSync_unlockstate(ModelSync *self)
+{
+	if(THMutex_unlock(self->state_mutex))
+		return 1;
+        return 0;
+}
+
+int  ModelSync_free(ModelSync *self)
+{
+	if (NULL != self && __sync_fetch_and_add(&self->refcount, -1) == 1)
+	{
+		free(self->model_mutex);	
+		free(self->state_mutex);
+		Status status;
+		CUDA_SAFE_SYNC_CALL(cudaFreeHost(self->free_data_), &status);
+		free(self);
+	}
+}
+
+int  ModelSync_initBuffer(ModelSync *self)
+{
+	if (NULL != self)
+	{
+		void *free_data = NULL, *data = NULL;
+		size_t size = self->dim_ * sizeof(float)+16;
+		Status status;
+		CUDA_SAFE_SYNC_CALL(cudaHostAlloc((void**) &free_data, size, cudaHostAllocPortable), &status);	
+		NERV_SET_STATUS(&status, NERV_NORMAL, 0);
+
+		data = (free_data ? (void *)( (((unsigned long)*(&free_data)) + 15) & ~0xFUL ) : NULL) ;
+		if (NULL != data)
+		{
+			self->data_ =  (float*)(data);
+			self->free_data_ = (float*)(free_data);
+		}
+		return 0;
+	}
+	return 1;
+}
+
+int  ModelSync_weightfromd(ModelSync *self, Matrix *dm)
+{
+
+	if (NULL != self && NULL != dm)
+	{
+		void *host_data_ = (void*)self->data_;
+		size_t width = dm->ncol * sizeof(float);
+		size_t src_pitch = dm->stride;
+		size_t dst_pitch = src_pitch;	
+		Status status;
+
+		CUDA_SAFE_SYNC_CALL(cudaMemcpy2D(host_data_+self->pos_, dst_pitch, dm->data.f, src_pitch, width, dm->nrow, cudaMemcpyDeviceToHost), &status);
+		NERV_SET_STATUS(&status, NERV_NORMAL, 0);
+		self->pos_ += dm->nrow * dm->stride;
+		return 0;
+	}
+	return 1;
+	
+}
+
+
+int  ModelSync_weighttod(ModelSync *self, Matrix *dm)
+{
+
+        if (NULL != self && NULL != dm) 
+        {   
+		void *host_data_ = (void*)self->data_;
+                size_t width = dm->ncol * sizeof(float);
+                size_t dst_pitch = dm->stride;
+                size_t src_pitch = dst_pitch;   
+		Status status;
+
+                CUDA_SAFE_SYNC_CALL(cudaMemcpy2D(dm->data.f, dst_pitch, host_data_+self->pos_, src_pitch, width, dm->nrow, cudaMemcpyHostToDevice), &status);
+		NERV_SET_STATUS(&status, NERV_NORMAL, 0);
+
+                self->pos_ += dm->nrow * dm->stride;
+		self->initialized_ = true;
+		return 0;
+        }   
+	return 1;
+}
+
+void ModelSync_syncinc(ModelSync *self)
+{
+	__sync_fetch_and_add(&self->threadcount, 1);
+}
+
+void ModelSync_syncdec(ModelSync *self)
+{
+	__sync_fetch_and_add(&self->threadcount, -1);
+}
+
+int ModelSync_threadcount(ModelSync *self)
+{
+	return self->threadcount;
+}
+
+/////////////////////////////////
+
+Xent* Xent_new()
+{
+	Xent *xent = (Xent*)malloc(sizeof(Xent));
+	memset(xent, 0, sizeof(Xent));
+	xent->refcount = 1;
+	return xent;
+}
+
+Xent* Xent_newWithId(long id)
+{
+	Xent *xent = (Xent*)id;
+	__sync_fetch_and_add(&xent->refcount, 1);
+	return xent;
+}
+
+Xent* Xent_newWithParm(size_t frames_, size_t correct_, double loss_, double entropy_)
+{
+	Xent *xent = (Xent*)malloc(sizeof(Xent));
+	xent->frames_ = frames_;
+	xent->correct_ = correct_;
+	xent->loss_ = loss_;
+	xent->entropy_ = entropy_;
+	xent->refcount = 1;
+	return xent;
+}
+
+long Xent_id(Xent *xent)
+{
+	return (long)(xent);
+}
+
+Xent* Xent_add(Xent *a, Xent *b)
+{
+	a->frames_ += b->frames_;
+	a->correct_ += b->correct_;
+	a->loss_ += b->loss_;
+	a->entropy_ += b->entropy_;
+	return a;
+}
+
+void Xent_free(Xent *xent)
+{	
+	if (NULL != xent && __sync_fetch_and_add(&xent->refcount, -1) == 1)
+	{ 
+		free(xent);
+		xent = NULL;
+	}
+}
+
+
+//////////////////////////////////
+
+Mse* Mse_new()
+{
+        Mse *mse = (Mse*)malloc(sizeof(Mse));
+        memset(mse, 0, sizeof(Mse));
+	mse->refcount = 1;
+        return mse;
+}
+
+Mse* Mse_newWithId(long id)
+{
+        Mse *mse = (Mse*)id;
+        __sync_fetch_and_add(&mse->refcount, 1);
+        return mse;
+}
+
+Mse* Mse_newWithParm(size_t frames_, double loss_)
+{
+        Mse *mse = (Mse*)malloc(sizeof(Mse));
+        mse->frames_ = frames_;
+        mse->loss_ = loss_;
+	mse->refcount = 1;
+        return mse;
+}
+
+
+long Mse_id(Mse *mse)
+{
+        return (long)(mse);
+}
+
+Mse* Mse_add(Mse *a, Mse *b)
+{
+        a->frames_ += b->frames_;
+        a->loss_ += b->loss_;
+        return a;
+}
+
+void Mse_free(Mse *mse)
+{
+        if (NULL != mse && __sync_fetch_and_add(&mse->refcount, -1) == 1)
+        {
+                free(mse);
+                mse = NULL;
+        }
+}
+
+//////////////////////////////////
+
+GlobalOption* GlobalOption_new()
+{
+        GlobalOption *option = (GlobalOption*)malloc(sizeof(GlobalOption));
+        option->refcount = 1;
+        return option;
+}
+
+GlobalOption* GlobalOption_newWithParm(int batch_size, float lrate, bool bp,const char *tr_scp, const char *cv_scp, const char *transf, const char *network)
+{
+        GlobalOption *option = (GlobalOption*)malloc(sizeof(GlobalOption));
+	option->batch_size = batch_size;
+	option->lrate = lrate;
+	option->bp = bp;
+	strncpy(option->tr_scp, tr_scp, strlen(tr_scp)+1);
+	strncpy(option->cv_scp, cv_scp, strlen(cv_scp)+1);
+	strncpy(option->transf, transf, strlen(transf)+1);
+	strncpy(option->network, network, strlen(network)+1);
+        option->refcount = 1;
+
+        return option;
+}
+
+GlobalOption* GlobalOption_newWithId(long id) 
+{
+        GlobalOption *option = (GlobalOption*)id;
+        __sync_fetch_and_add(&option->refcount, 1); 
+        return option;
+}
+
+
+
+long GlobalOption_id(GlobalOption *option)
+{
+        return (long)(option);
+}
+
+void GlobalOption_free(GlobalOption *option)
+{
+        if (NULL != option && __sync_fetch_and_add(&option->refcount, -1) == 1)
+        {   
+                free(option);
+                option = NULL;
+        }   
+}
+
+
diff --git a/fastnn/threads/Makefile b/fastnn/threads/Makefile
index 17958f9..4205adc 100644
--- a/fastnn/threads/Makefile
+++ b/fastnn/threads/Makefile
@@ -35,6 +35,7 @@ $(OBJ_DIR)/%.o: %.c $(patsubst /%.o,/%.c,$@)
 
 $(LIBS): $(OBJS)
 	gcc -shared -o $@ $^ $(LDFLAGS)  -Wl,-rpath=$(LIB_PATH) -L$(LIB_PATH) -lnervcore -lluaT -lpthread
+	cp $@ $(LIB_PATH)/
 
 clean:
 	-rm -rf $(OBJ_DIR)
diff --git a/nerv/io/sgd_buffer.lua b/nerv/io/sgd_buffer.lua
index 3f854f0..65d6da1 100644
--- a/nerv/io/sgd_buffer.lua
+++ b/nerv/io/sgd_buffer.lua
@@ -57,7 +57,7 @@ function SGDBuffer:saturate()
                 buff.data:copy_from(buff.leftover, 0, lrow)
                 buff.leftover = nil
             end
-            nerv.info("buffer leftover: %d\n", lrow)
+            nerv.info("buffer leftover: %d", lrow)
             reader.tail = lrow
             reader.has_leftover = false
         end
@@ -107,7 +107,7 @@ function SGDBuffer:get_data()
         if not self:saturate() then
             return nil -- the remaining data cannot build a batch
         end
-        nerv.info("%.3fs to fill the buffer", os.clock() - t)
+        --nerv.info("%.3fs to fill the buffer", os.clock() - t)
     end
     if self.head + batch_size > self.tail then
         return nil -- the remaining data cannot build a batch
diff --git a/nerv/lib/common.h b/nerv/lib/common.h
index 6878e34..a4e3582 100644
--- a/nerv/lib/common.h
+++ b/nerv/lib/common.h
@@ -59,6 +59,8 @@ typedef struct Status {
             nerv_error_status(L, &status); \
     } while (0)
 
+#define PROFILE_HASHMAP_SIZE 123457
+
 typedef struct HashNode {
     const char *key;
     void *val;
diff --git a/nerv/lib/matrix/cumatrix.c b/nerv/lib/matrix/cumatrix.c
index a5991ab..c913db2 100644
--- a/nerv/lib/matrix/cumatrix.c
+++ b/nerv/lib/matrix/cumatrix.c
@@ -2,7 +2,7 @@
 #include "../common.h"
 #include "cuda_helper.h"
 #include <string.h>
-#define PROFILE_HASHMAP_SIZE 123457
+
 static cublasHandle_t cublas_handle;
 static cudaEvent_t profile_start, profile_stop;
 static HashMap *profile;
diff --git a/nerv/nn/layer_repo.lua b/nerv/nn/layer_repo.lua
index ef333a7..8473727 100644
--- a/nerv/nn/layer_repo.lua
+++ b/nerv/nn/layer_repo.lua
@@ -13,7 +13,7 @@ function LayerRepo:add_layers(layer_spec, param_repo, global_conf)
             if layers[id] ~= nil then
                 nerv.error("a layer with id %s already exists", id)
             end
-            nerv.info("create layer: %s", id)
+            --nerv.info("create layer: %s", id)
             if type(spec[2]) ~= "table" then
                 nerv.error("layer config table is need")
             end