release layer wgs and version update

include/tkDNN/Layer.h

@@ -108,24 +108,64 @@ public:
 
     // additional bias for DCN
     bool additional_bias;
-    dnnType *bias2_h, *bias2_d;
+    dnnType *bias2_h = nullptr, *bias2_d = nullptr;
 
     //batchnorm
     bool batchnorm;
-    dnnType *power_h;
-    dnnType *scales_h,   *scales_d;
-    dnnType *mean_h,     *mean_d;
-    dnnType *variance_h, *variance_d;
+    dnnType *power_h    = nullptr;
+    dnnType *scales_h   = nullptr,   *scales_d = nullptr;
+    dnnType *mean_h     = nullptr,     *mean_d = nullptr;
+    dnnType *variance_h = nullptr, *variance_d = nullptr;
 
     //fp16
-    __half *data16_h, *bias16_h;
-    __half *data16_d, *bias16_d;
-    __half *bias216_h, *bias216_d;
+    __half *data16_h  = nullptr, *bias16_h  = nullptr;
+    __half *data16_d  = nullptr, *bias16_d  = nullptr;
+    __half *bias216_h = nullptr, *bias216_d = nullptr;
 
-    __half *power16_h,    *power16_d;
-    __half *scales16_h,   *scales16_d;
-    __half *mean16_h,     *mean16_d;
-    __half *variance16_h, *variance16_d;
+    __half *power16_h    = nullptr;
+    __half *scales16_h   = nullptr,   *scales16_d = nullptr;
+    __half *mean16_h     = nullptr,     *mean16_d = nullptr;
+    __half *variance16_h = nullptr, *variance16_d = nullptr;
+
+    void releaseHost(bool release32 = true, bool release16 = true) {
+        if(release32) {
+            if(    data_h != nullptr) { delete []     data_h;     data_h = nullptr; }
+            if(    bias_h != nullptr) { delete []     bias_h;     bias_h = nullptr; }
+            if(   bias2_h != nullptr) { delete []    bias2_h;    bias2_h = nullptr; }
+            if(  scales_h != nullptr) { delete []   scales_h;   scales_h = nullptr; }
+            if(    mean_h != nullptr) { delete []     mean_h;     mean_h = nullptr; }
+            if(variance_h != nullptr) { delete [] variance_h; variance_h = nullptr; }
+            if(   power_h != nullptr) { delete []    power_h;    power_h = nullptr; }
+        }
+        if(net->fp16 && release16) {
+            if(    data16_h != nullptr) { delete []     data16_h;     data16_h = nullptr; }
+            if(    bias16_h != nullptr) { delete []     bias16_h;     bias16_h = nullptr; }
+            if(   bias216_h != nullptr) { delete []    bias216_h;    bias216_h = nullptr; }
+            if(  scales16_h != nullptr) { delete []   scales16_h;   scales16_h = nullptr; }
+            if(    mean16_h != nullptr) { delete []     mean16_h;     mean16_h = nullptr; }
+            if(variance16_h != nullptr) { delete [] variance16_h; variance16_h = nullptr; } 
+            if(   power16_h != nullptr) { delete []    power16_h;    power16_h = nullptr; }
+
+        }
+    }
+    void releaseDevice(bool release32 = true, bool release16 = true) {
+        if(release32) {
+            if(    data_d != nullptr) { cudaFree(    data_d);     data_d = nullptr; }
+            if(    bias_d != nullptr) { cudaFree(    bias_d);     bias_d = nullptr; }
+            if(   bias2_d != nullptr) { cudaFree(   bias2_d);    bias2_d = nullptr; }
+            if(  scales_d != nullptr) { cudaFree(  scales_d);   scales_d = nullptr; }
+            if(    mean_d != nullptr) { cudaFree(    mean_d);     mean_d = nullptr; }
+            if(variance_d != nullptr) { cudaFree(variance_d); variance_d = nullptr; }
+        }
+        if(net->fp16 && release16) {
+            if(    data16_d != nullptr) { cudaFree(    data16_d);     data16_d = nullptr; }
+            if(    bias16_d != nullptr) { cudaFree(    bias16_d);     bias16_d = nullptr; }
+            if(   bias216_d != nullptr) { cudaFree(   bias216_d);    bias216_d = nullptr; }
+            if(  scales16_d != nullptr) { cudaFree(  scales16_d);   scales16_d = nullptr; }
+            if(    mean16_d != nullptr) { cudaFree(    mean16_d);     mean16_d = nullptr; }
+            if(variance16_d != nullptr) { cudaFree(variance16_d); variance16_d = nullptr; } 
+        }
+    }
 };
 
 

include/tkDNN/tkdnn.h

@@ -5,4 +5,4 @@
 #include "Layer.h"
 #include "NetworkRT.h"
 
-#define TKDNN_VERSION 400
+#define TKDNN_VERSION 500

src/LayerWgs.cpp

@@ -80,7 +80,7 @@ LayerWgs::LayerWgs(Network *net, int inputs, int outputs,
         variance16_h = new __half[b_size];
         scales16_h   = new __half[b_size];
 
-        cudaMalloc(&power16_d, b_size*sizeof(__half));
+        //cudaMalloc(&power16_d, b_size*sizeof(__half));
         cudaMalloc(&mean16_d, b_size*sizeof(__half));
         cudaMalloc(&variance16_d, b_size*sizeof(__half));
         cudaMalloc(&scales16_d, b_size*sizeof(__half));
@@ -91,11 +91,10 @@ LayerWgs::LayerWgs(Network *net, int inputs, int outputs,
 
         //init power array of ones
         cudaMemcpy(tmp_d, power_h, b_size*sizeof(float), cudaMemcpyHostToDevice);
-        float2half(tmp_d, power16_d, b_size);
-        cudaMemcpy(power16_h, power16_d, b_size*sizeof(__half), cudaMemcpyDeviceToHost);
+        //float2half(tmp_d, power16_d, b_size);
+        //cudaMemcpy(power16_h, power16_d, b_size*sizeof(__half), cudaMemcpyDeviceToHost);
 
         //mean array
-
         cudaMemcpy(tmp_d, mean_h, b_size*sizeof(float), cudaMemcpyHostToDevice);
         float2half(tmp_d, mean16_d, b_size);
         cudaMemcpy(mean16_h, mean16_d, b_size*sizeof(__half), cudaMemcpyDeviceToHost);
@@ -109,24 +108,14 @@ LayerWgs::LayerWgs(Network *net, int inputs, int outputs,
         //conver scales
         float2half(scales_d, scales16_d, b_size);
         cudaMemcpy(scales16_h, scales16_d, b_size*sizeof(__half), cudaMemcpyDeviceToHost);
+    
+        cudaFree(tmp_d);
     }
 }
 
 LayerWgs::~LayerWgs() {
-
-    delete [] data_h;
-    delete [] bias_h;
-    checkCuda( cudaFree(data_d) );
-    checkCuda( cudaFree(bias_d) );
-
-    if(batchnorm) {
-        delete [] scales_h;
-        delete [] mean_h;
-        delete [] variance_h;
-        checkCuda( cudaFree(scales_d) );
-        checkCuda( cudaFree(mean_d) );
-        checkCuda( cudaFree(variance_d) );
-    }
+    releaseHost();
+    releaseDevice();
 }
 
 }}