Make code safer and simpler.

theano/gpuarray/c_code/dnn_conv_base.c

@@ -59,7 +59,6 @@ if (APPLY_SPECIFIC(kerns) != NULL)
 
 #section support_code
 #include <sstream>
-#include <vector>
 #include <string>
 #if __cplusplus < 201103L
 #include <tr1/unordered_map>
@@ -70,19 +69,16 @@ typedef std::unordered_map<std::string, AlgoRec> AlgoCache;
 #endif
 #include "pthread.h"
 
-#line 69 "dnn_conv_base.c"
-
-using std::vector;
-using std::string;
+#line 73 "dnn_conv_base.c"
 
 pthread_mutex_t  algoMutex;
 AlgoCache        algoCache;
 
-static cudnnStatus_t checkCudnnStatus(cudnnStatus_t err)
+static cudnnStatus_t checkCudnnStatus(cudnnStatus_t err, const char* msg)
 {
     if (err != CUDNN_STATUS_SUCCESS) {
-        PyErr_Format(PyExc_RuntimeError, "CUDNN Error: %s",
-                     cudnnGetErrorString(err));
+        PyErr_Format(PyExc_RuntimeError, "CUDNN Error: %s: %s",
+                     msg, cudnnGetErrorString(err));
     }
     return err;
 }
@@ -105,64 +101,69 @@ c_get_largest_free_block_size(PyGpuContextObject *c)
 
 static std::string shape(int* res, int size)
 {
-    std::stringstream s;
-    if (size>0) {
+    std::ostringstream s;
+    if (size > 0) {
       
-      s<<res[0];
-      for (int i=1; i< size; ++i)
+      s << res[0];
+      for (int i = 1; i < size; ++i)
         s <<',' << res[i];
     }
-    return std::string(s.str().c_str());
+    return s.str();
 }
 
-
 static std::string shape(cudnnTensorDescriptor_t t)
 {
-    std::vector<int> res;
-    std::vector<int> stride;
-        
+    // cuDNN can handle up to CUDNN_DIM_MAX dimensions.
+    int res[CUDNN_DIM_MAX];
+    int stride[CUDNN_DIM_MAX];
     int nbDims;
     cudnnDataType_t type;
-    checkCudnnStatus(cudnnGetTensorNdDescriptor(t, 0, &type, &nbDims,0,0));
-    res.resize(nbDims);
-    stride.resize(nbDims);
-    checkCudnnStatus(cudnnGetTensorNdDescriptor(t, nbDims, &type, &nbDims, res.data(), stride.data()));
-    return shape(&res[0], nbDims) + shape(&stride[0], nbDims);
-    
+    checkCudnnStatus(cudnnGetTensorNdDescriptor(t, CUDNN_DIM_MAX, &type, &nbDims, res, stride),
+                     "error getting tensor description");
+    if (PyErr_Occurred()) return "";
+    return shape(res, nbDims) + "," + shape(stride, nbDims);
 };
 
 static std::string shape(cudnnFilterDescriptor_t t, cudnnDataType_t* type)
 {
     cudnnTensorFormat_t format;
-    int sizes = 8;
-    
-    std::vector<int> res(sizes);
+    int res[CUDNN_DIM_MAX];
     int outDims;
-    checkCudnnStatus(cudnnGetFilterNdDescriptor(t, sizes, type, &format, &outDims, res.data()));
-    return shape(&res[0], outDims);
+    checkCudnnStatus(cudnnGetFilterNdDescriptor(t, CUDNN_DIM_MAX, type, &format, &outDims, res),
+                     "error getting filter description");
+    if (PyErr_Occurred()) return "";
+    return shape(res, outDims);
 };
 
 static std::string shape(cudnnConvolutionDescriptor_t convDesc)
 {
-    const int maxDim = 5;
-    int nDim=0;
+    int nDim;
     cudnnConvolutionMode_t mode;
     cudnnDataType_t        computeType;
     
-    int                                 padA[maxDim];
-    int                                 strideA[maxDim];
-    int                                 dilationA[maxDim];    
+    int                                 padA[5];
+    int                                 strideA[5];
+    int                                 dilationA[5];
 
     checkCudnnStatus(
-        cudnnGetConvolutionNdDescriptor( convDesc, maxDim,
+        cudnnGetConvolutionNdDescriptor( convDesc, 5,
                                          &nDim,
                                          &padA[0],
                                          &strideA[0],
                                          &dilationA[0],
                                          &mode,
-                                         &computeType ));
-    
-    return std::string("-mode ") + (((int)mode==0) ? "conv" : "corr") + " -padA" + shape(padA,nDim) + " -convStrideA " + shape(strideA, nDim)  + " -dilationA " + shape(dilationA, nDim);
+                                         &computeType ),
+        "error getting convolution description");
+    if (PyErr_Occurred()) return "";
+    
+    return (std::string("-mode ") +
+            ((mode == CUDNN_CONVOLUTION) ? "conv" : "cross") +
+            " -pad " +
+            shape(padA, nDim) +
+            " -subsample " +
+            shape(strideA, nDim) +
+            " -dilation " +
+            shape(dilationA, nDim));
 }
 
 static bool all_aligned(cudnnDataType_t type, void* in, void* out, void* filter)
@@ -182,7 +183,7 @@ static std::string dnn_conv_shape(cudnnTensorDescriptor_t inputDesc, PyGpuArrayO
 				  PyGpuArrayObject* output, int groups)
 {
     cudnnDataType_t  dType;
-    std::stringstream s;
+    std::ostringstream s;
     int expected_output_dims[5] = {0};
     cudnnStatus_t err = cudnnGetConvolutionNdForwardOutputDim(convDesc, inputDesc, filterDesc,
 							      PyGpuArray_NDIM(filter), expected_output_dims);
@@ -221,16 +222,20 @@ static std::string dnn_conv_shape(cudnnTensorDescriptor_t inputDesc, PyGpuArrayO
         return "";
       }
     }
-    
-    s << "-g" << groups << " -dimA" << shape(inputDesc) << " -filtA" <<
-      shape(filterDesc, &dType) << shape(convDesc);    
+    std::string shapeInput = shape(inputDesc);
+    std::string shapeFilter = shape(filterDesc, &dType);
+    std::string shapeConvDesc = shape(convDesc);
+    if (shapeInput.empty() || shapeFilter.empty() || shapeConvDesc.empty())
+        return "";
+    s << "-g " << groups << " -dim " << shapeInput << " -filt " <<
+      shapeFilter << " " << shapeConvDesc;
     
 // there have to be entries for both aligned and not
     if (!all_aligned(dType, PyGpuArray_DEV_DATA(input), PyGpuArray_DEV_DATA(output), PyGpuArray_DEV_DATA(filter)))
     {
-      s << " [unaligned] ";
+      s << " [unaligned]";
     }
-    return std::string(s.str().c_str());
+    return s.str();
 }
 
 static void dnn_conv_update_cache(const std::string& hash, const AlgoRec& rec)
@@ -240,15 +245,11 @@ static void dnn_conv_update_cache(const std::string& hash, const AlgoRec& rec)
   pthread_mutex_unlock(&algoMutex);
 }
 
-
 static const AlgoRec* dnn_conv_check_cache(const std::string& hash)
 {
   pthread_mutex_lock(&algoMutex);    
-  bool cacheHit = false;
   const AlgoRec* ret = 0;
   
-  // cout << "dnn_conv_check_cache: "<< hash << endl;
-  
   AlgoCache::iterator hit = algoCache.find(hash);
   
   if (hit != algoCache.end())
@@ -257,4 +258,3 @@ static const AlgoRec* dnn_conv_check_cache(const std::string& hash)
   pthread_mutex_unlock(&algoMutex);
   return ret;
 }
-

theano/gpuarray/c_code/dnn_fwd.c

 #section init_code_struct
 reuse_algo = 0;
+use_cached = 0;
 prev_algo.algo = PARAMS->conv_algo;
 prev_algo.mathType = CUDNN_DEFAULT_MATH;
 prev_algo.dataType = CUDNN_DATA_FLOAT;
-hash_prefix = std::string("FW| GPU#");
+hash_prefix = std::string("FWD|GPU#");
 
 #section support_code_struct
-#line 12 "dnn_fwd.c"
+#line 11 "dnn_fwd.c"
 int     reuse_algo;
 bool    use_cached;
 AlgoRec prev_algo;
@@ -97,6 +98,7 @@ APPLY_SPECIFIC(conv_fwd)(PyGpuArrayObject *input, PyGpuArrayObject *kerns,
   #endif
 
   size_t free = c_get_largest_free_block_size(c);
+  if (PyErr_Occurred()) return 1;
 
   cuda_enter(c->ctx);
 
@@ -108,7 +110,7 @@ APPLY_SPECIFIC(conv_fwd)(PyGpuArrayObject *input, PyGpuArrayObject *kerns,
       hashkey = dnn_conv_shape(APPLY_SPECIFIC(input), input, APPLY_SPECIFIC(kerns), kerns, desc, *output, groups);
       if (hashkey.empty())
         return 1;
-      hashkey =  hash_prefix + pci_id + hashkey;
+      hashkey =  hash_prefix + pci_id + " " + hashkey;
       // check out cache
       const AlgoRec* cached = dnn_conv_check_cache(hashkey);
       if (cached) {
@@ -142,9 +144,6 @@ APPLY_SPECIFIC(conv_fwd)(PyGpuArrayObject *input, PyGpuArrayObject *kerns,
           free);
         gpudata_release(tmpmem);
 
-	//	fprintf(stderr, "(cudnnFindConvolutionForwardAlgorithmEx: (err:%d), algo: %d, mem: %ld, free: %ld\n",
-	//	err, choice.algo, choice.memory, free);
-	
         if (err != CUDNN_STATUS_SUCCESS) {
           PyErr_Format(PyExc_RuntimeError,
                        "error selecting convolution algo: %s",
@@ -152,14 +151,6 @@ APPLY_SPECIFIC(conv_fwd)(PyGpuArrayObject *input, PyGpuArrayObject *kerns,
           cuda_exit(c->ctx);
           return 1;
         }
-        algo = choice.algo;
-        prev_algo.algo = (int)algo;
-        prev_algo.wsSize = worksize = choice.memory;
-#if CUDNN_MAJOR >= 7        
-        prev_algo.mathType = mathtype = choice.mathType;
-#endif
-        // Add to the cache
-        dnn_conv_update_cache(hashkey, prev_algo);
 
         #ifdef DEBUG
         if (count == 0) {
@@ -173,6 +164,17 @@ APPLY_SPECIFIC(conv_fwd)(PyGpuArrayObject *input, PyGpuArrayObject *kerns,
         } // Else, count is necessarly 1 for current implementation.
         #endif
 
+        algo = choice.algo;
+        prev_algo.algo = (int)algo;
+        prev_algo.wsSize = worksize = choice.memory;
+#if CUDNN_MAJOR >= 7
+        prev_algo.mathType = mathtype = choice.mathType;
+#endif
+        // Add to the cache
+        dnn_conv_update_cache(hashkey, prev_algo);
+        // NB: It is added again later to cqche,
+        // so maybe this line could be removed.
+
       } else {
         err = cudnnGetConvolutionForwardAlgorithm(
           params->handle, APPLY_SPECIFIC(input), APPLY_SPECIFIC(kerns),
@@ -188,7 +190,6 @@ APPLY_SPECIFIC(conv_fwd)(PyGpuArrayObject *input, PyGpuArrayObject *kerns,
         prev_algo.algo = algo;
         // no tensor_op returned from Get()
         prev_algo.mathType = mathtype = CUDNN_DEFAULT_MATH;
-	  // fprintf(stderr, "(cudnnGetConvolutionForwardAlgorithm: (err:%d), algo: %d\n", err, algo);
       }
     }
   }
@@ -233,7 +234,8 @@ APPLY_SPECIFIC(conv_fwd)(PyGpuArrayObject *input, PyGpuArrayObject *kerns,
   // save worksize for next time/cache
   prev_algo.wsSize = worksize;
 
-  // Add to the cache
+  // Add to the cache, even if this node use *_once algo
+  // (in case the user specify the algo per layer and not globally).
   if (params->choose_algo)
     dnn_conv_update_cache(hashkey, prev_algo);
 
@@ -241,13 +243,14 @@ APPLY_SPECIFIC(conv_fwd)(PyGpuArrayObject *input, PyGpuArrayObject *kerns,
   if (params->choose_algo) {
     if (0 != theano_enum_to_string_cudnnConvolutionFwdAlgo_t(algo, algorithm_name))
       return 1;
-    fprintf(stderr, "%s%s algo: %d %s%s ws: %ld, tensor: %d hash:%s\n",
-            params->choose_algo ? "[A]": "" ,
-            params->choose_time ? "[T]": "" ,
-            algo, // algorithm_name,
+    fprintf(stderr, "(using %s %s%s%s%s, ws:%ld, hash:%s)\n",
+            algorithm_name,
+            params->choose_time ? "(timed)": "" ,
             reuse_algo ? "(reused)" : "",
             use_cached ? "(cache)": "",
-            worksize, mathtype, hashkey.c_str()
+            mathtype == CUDNN_TENSOR_OP_MATH ? "(tensor op)" : "",
+            worksize,
+            hashkey.c_str()
       );
   }
 #endif

theano/gpuarray/c_code/dnn_gi.c

@@ -3,10 +3,11 @@ prev_algo.algo = PARAMS->conv_algo;
 prev_algo.mathType = CUDNN_DEFAULT_MATH;
 prev_algo.dataType = CUDNN_DATA_FLOAT;
 reuse_algo = 0;
-hash_prefix = std::string("GI| GPU#");
-#section support_code_struct
+use_cached = 0;
+hash_prefix = std::string("GI|GPU#");
 
-#line 12 "dnn_gi.c"
+#section support_code_struct
+#line 11 "dnn_gi.c"
 int     reuse_algo;
 bool    use_cached;
 AlgoRec prev_algo;
@@ -102,7 +103,7 @@ APPLY_SPECIFIC(conv_gi)(PyGpuArrayObject *kerns, PyGpuArrayObject *output,
     hashkey=dnn_conv_shape(APPLY_SPECIFIC(input), *input, APPLY_SPECIFIC(kerns), kerns, desc, output, groups);
     if (hashkey.empty())
       return 1;
-    hashkey = hash_prefix + pci_id + hashkey;
+    hashkey = hash_prefix + pci_id + " " + hashkey;
     const AlgoRec* cached = dnn_conv_check_cache(hashkey);
     if (cached) {
       prev_algo = *cached;
@@ -111,6 +112,7 @@ APPLY_SPECIFIC(conv_gi)(PyGpuArrayObject *kerns, PyGpuArrayObject *output,
   }
 
   size_t free = c_get_largest_free_block_size(c);
+  if (PyErr_Occurred()) return 1;
 
   cuda_enter(c->ctx);
 
@@ -140,15 +142,6 @@ APPLY_SPECIFIC(conv_gi)(PyGpuArrayObject *kerns, PyGpuArrayObject *output,
         return 1;
       }
 
-      algo = choice.algo;
-        prev_algo.algo = (int)algo;
-        prev_algo.wsSize = worksize = choice.memory;
-#if CUDNN_MAJOR >= 7
-        prev_algo.mathType = mathtype = choice.mathType;
-#endif
-        // Add to the cache
-	dnn_conv_update_cache(hashkey, prev_algo);
-
       #ifdef DEBUG
       if (count == 0) {
           PyErr_SetString(PyExc_RuntimeError, "No best-timed conv gradinput algorithm found");
@@ -161,6 +154,15 @@ APPLY_SPECIFIC(conv_gi)(PyGpuArrayObject *kerns, PyGpuArrayObject *output,
       } // Else, count is necessarly 1 for current implementation.
       #endif
 
+      algo = choice.algo;
+      prev_algo.algo = (int)algo;
+      prev_algo.wsSize = worksize = choice.memory;
+#if CUDNN_MAJOR >= 7
+      prev_algo.mathType = mathtype = choice.mathType;
+#endif
+    // Add to the cache
+	dnn_conv_update_cache(hashkey, prev_algo);
+
     } else {
       err = cudnnGetConvolutionBackwardDataAlgorithm(
         params->handle, APPLY_SPECIFIC(kerns), APPLY_SPECIFIC(output),
@@ -258,13 +260,14 @@ APPLY_SPECIFIC(conv_gi)(PyGpuArrayObject *kerns, PyGpuArrayObject *output,
     if (0 != theano_enum_to_string_cudnnConvolutionBwdDataAlgo_t(algo, algorithm_name))
         return 1;
     // NB: This is printed only when algorithm is chosen at runtime.
-    fprintf(stderr, "%s%s algo: %d %s%s ws: %ld, tensor: %d hash:%s\n",
-            params->choose_algo ? "[A]": "" ,
-            params->choose_time ? "[T]": "" ,
-            algo, // algorithm_name,
+    fprintf(stderr, "(using %s %s%s%s%s, ws:%ld, hash:%s)\n",
+            algorithm_name,
+            params->choose_time ? "(timed)": "" ,
             reuse_algo ? "(reused)" : "",
             use_cached ? "(cache)": "",
-            worksize, mathtype, hashkey.c_str()
+            mathtype == CUDNN_TENSOR_OP_MATH ? "(tensor op)" : "",
+            worksize,
+            hashkey.c_str()
       );
   }
 #endif

theano/gpuarray/c_code/dnn_gw.c

@@ -3,11 +3,11 @@ prev_algo.algo = PARAMS->conv_algo;
 prev_algo.mathType = CUDNN_DEFAULT_MATH;
 prev_algo.dataType = CUDNN_DATA_FLOAT;
 reuse_algo = 0;
-hash_prefix = std::string("GW| GPU#");
+use_cached = 0;
+hash_prefix = std::string("GW|GPU#");
 
 #section support_code_struct
 #line 11 "dnn_gw.c"
-
 int     reuse_algo;
 bool    use_cached;
 AlgoRec prev_algo;
@@ -96,6 +96,7 @@ APPLY_SPECIFIC(conv_gw)(PyGpuArrayObject *input, PyGpuArrayObject *output,
   std::string hashkey ;
 
   size_t free = c_get_largest_free_block_size(c);
+  if (PyErr_Occurred()) return 1;
 
   cuda_enter(c->ctx);
 
@@ -148,15 +149,6 @@ APPLY_SPECIFIC(conv_gw)(PyGpuArrayObject *input, PyGpuArrayObject *output,
           return 1;
         }
 
-        algo = choice.algo;
-        prev_algo.algo = (int)algo;
-        prev_algo.wsSize = worksize = choice.memory;
-#if CUDNN_MAJOR >= 7        
-        prev_algo.mathType = mathtype = choice.mathType;
-#endif
-        // Add to the cache
-        dnn_conv_update_cache(hashkey, prev_algo);
-
         #ifdef DEBUG
         if (count == 0) {
             PyErr_SetString(PyExc_RuntimeError, "No best-timed conv gradweight algorithm found");
@@ -169,6 +161,15 @@ APPLY_SPECIFIC(conv_gw)(PyGpuArrayObject *input, PyGpuArrayObject *output,
         } // Else, count is necessarly 1 for current implementation.
         #endif
 
+        algo = choice.algo;
+        prev_algo.algo = (int)algo;
+        prev_algo.wsSize = worksize = choice.memory;
+#if CUDNN_MAJOR >= 7
+        prev_algo.mathType = mathtype = choice.mathType;
+#endif
+        // Add to the cache
+        dnn_conv_update_cache(hashkey, prev_algo);
+
       } else {
         err = cudnnGetConvolutionBackwardFilterAlgorithm(
           params->handle, APPLY_SPECIFIC(input), APPLY_SPECIFIC(output),
@@ -231,13 +232,14 @@ APPLY_SPECIFIC(conv_gw)(PyGpuArrayObject *input, PyGpuArrayObject *output,
     if (0 != theano_enum_to_string_cudnnConvolutionBwdFilterAlgo_t(algo, algorithm_name))
       return 1;
     // NB: This is printed only when algorithm is chosen at runtime.
-    fprintf(stderr, "%s%s algo: %d %s%s ws: %ld, tensor: %d hash:%s\n",
-            params->choose_algo ? "[A]": "" ,
-            params->choose_time ? "[T]": "" ,
-            algo, // algorithm_name,
+    fprintf(stderr, "(using %s %s%s%s%s, ws:%ld, hash:%s)\n",
+            algorithm_name,
+            params->choose_time ? "(timed)": "" ,
             reuse_algo ? "(reused)" : "",
             use_cached ? "(cache)": "",
-            worksize, mathtype, hashkey.c_str()
+            mathtype == CUDNN_TENSOR_OP_MATH ? "(tensor op)" : "",
+            worksize,
+            hashkey.c_str()
      );
   }
 #endif

theano/gpuarray/dnn.py

@@ -399,7 +399,7 @@ class DnnBase(COp):
         return []
 
     def c_code_cache_version(self):
-        return (super(DnnBase, self).c_code_cache_version(), version(), 3)
+        return (super(DnnBase, self).c_code_cache_version(), version(), 1)
 
 
 class GpuDnnConvDesc(COp):