Merge branch 'master' into batchwise_dropout

PyTorch/setup.py

@@ -35,6 +35,7 @@ if torch.cuda.is_available():
             'sparseconvnet/SCN/header_cpu.h',
             'sparseconvnet/SCN/header_gpu.h'],
         sources=[],
+        include_dirs=[os.path.expandvars('$CUDA_HOME') + '/include'],
         extra_objects=[
             this_dir +
             '/sparseconvnet/SCN/init.cu.o'],

PyTorch/sparseconvnet/SCN/generic/CPU/BatchNormalization.cpp

@@ -14,18 +14,20 @@ extern "C" void scn_R_(BatchNormalization_updateOutput)(
     THTensor *saveInvStd, THTensor *runningMean, THTensor *runningVar,
     THTensor *weight, THTensor *bias, real eps, real momentum, bool train,
     real leakiness) {
-
   THTensor_(resizeAs)(output_features, input_features);
+  if (input_features->nDimension == 2) {
     auto nActive = input_features->size[0];
     auto nPlanes = input_features->size[1];
     auto input_stride = input_features->stride[0];
     auto output_stride = output_features->stride[0];
     BatchNormalization_ForwardPass<real>(
         THTensor_(data)(input_features), THTensor_(data)(output_features),
-      nPlanes, input_stride, output_stride, nActive, THTensor_(data)(saveMean),
-      THTensor_(data)(saveInvStd), THTensor_(data)(runningMean),
-      THTensor_(data)(runningVar), THOptionalTensorData(weight),
-      THOptionalTensorData(bias), eps, momentum, train, leakiness);
+        nPlanes, input_stride, output_stride, nActive,
+        THTensor_(data)(saveMean), THTensor_(data)(saveInvStd),
+        THTensor_(data)(runningMean), THTensor_(data)(runningVar),
+        THOptionalTensorData(weight), THOptionalTensorData(bias), eps, momentum,
+        train, leakiness);
+  }
 }
 
 extern "C" void scn_R_(BatchNormalizationInTensor_updateOutput)(
@@ -34,6 +36,7 @@ extern "C" void scn_R_(BatchNormalizationInTensor_updateOutput)(
     THTensor *weight, THTensor *bias, real eps, real momentum, bool train,
     real leakiness) {
 
+  if (input_features->nDimension == 2) {
     auto nActive = input_features->size[0];
     auto nPlanes = input_features->size[1];
     auto input_stride = input_features->stride[0];
@@ -41,10 +44,12 @@ extern "C" void scn_R_(BatchNormalizationInTensor_updateOutput)(
 
     BatchNormalization_ForwardPass<real>(
         THTensor_(data)(input_features), THTensor_(data)(output_features),
-      nPlanes, input_stride, output_stride, nActive, THTensor_(data)(saveMean),
-      THTensor_(data)(saveInvStd), THTensor_(data)(runningMean),
-      THTensor_(data)(runningVar), THOptionalTensorData(weight),
-      THOptionalTensorData(bias), eps, momentum, train, leakiness);
+        nPlanes, input_stride, output_stride, nActive,
+        THTensor_(data)(saveMean), THTensor_(data)(saveInvStd),
+        THTensor_(data)(runningMean), THTensor_(data)(runningVar),
+        THOptionalTensorData(weight), THOptionalTensorData(bias), eps, momentum,
+        train, leakiness);
+  }
 }
 
 extern "C" void scn_R_(BatchNormalization_backward)(
@@ -55,6 +60,7 @@ extern "C" void scn_R_(BatchNormalization_backward)(
     real leakiness) {
 
   THTensor_(resizeAs)(d_input_features, input_features);
+  if (input_features->nDimension == 2) {
     auto nActive = input_features->size[0];
     auto nPlanes = input_features->size[1];
     auto input_stride = input_features->stride[0];
@@ -62,10 +68,12 @@ extern "C" void scn_R_(BatchNormalization_backward)(
     BatchNormalization_BackwardPass<real>(
         THTensor_(data)(input_features), THTensor_(data)(d_input_features),
         THTensor_(data)(output_features), THTensor_(data)(d_output_features),
-      nPlanes, input_stride, output_stride, nActive, THTensor_(data)(saveMean),
-      THTensor_(data)(saveInvStd), THTensor_(data)(runningMean),
-      THTensor_(data)(runningVar), THOptionalTensorData(weight),
-      THOptionalTensorData(bias), THOptionalTensorData(d_weight),
-      THOptionalTensorData(d_bias), leakiness);
+        nPlanes, input_stride, output_stride, nActive,
+        THTensor_(data)(saveMean), THTensor_(data)(saveInvStd),
+        THTensor_(data)(runningMean), THTensor_(data)(runningVar),
+        THOptionalTensorData(weight), THOptionalTensorData(bias),
+        THOptionalTensorData(d_weight), THOptionalTensorData(d_bias),
+        leakiness);
+  }
 }
 #endif

PyTorch/sparseconvnet/SCN/generic/CPU/Convolution.cpp

@@ -23,6 +23,8 @@ extern "C" double scn_DR_(Convolution_updateOutput)(
   if (not bias)
     THTensor_(zero)(output_features);
 
+  double flops = 0;
+  if (nActive) {
     auto iF = THTensor_(data)(input_features);
     auto oF = THTensor_(data)(output_features);
     auto ip = input_features->size[1];
@@ -31,9 +33,9 @@ extern "C" double scn_DR_(Convolution_updateOutput)(
     auto b = THOptionalTensorData(bias);
     Convolution_ForwardPass(iF, ip, ip, oF, op, op, w, b, _rules, nActive,
                             THBlas_(gemm));
-  double flops = 0;
     for (auto &r : _rules)
       flops += r.size() / 2 * ip * op;
+  }
   return flops;
 }
 
@@ -51,6 +53,7 @@ extern "C" void scn_DR_(Convolution_backward)(
   THTensor_(resizeAs)(d_input_features, input_features);
   THTensor_(zero)(d_input_features);
 
+  if (nActive) {
     auto iF = THTensor_(data)(input_features);
     auto diF = THTensor_(data)(d_input_features);
     auto doF = THTensor_(data)(d_output_features);
@@ -62,6 +65,7 @@ extern "C" void scn_DR_(Convolution_backward)(
 
     Convolution_BackwardPass(iF, diF, ip, ip, doF, op, op, w, dw, db, _rules,
                              nActive, THBlas_(gemm));
+  }
 }
 
 extern "C" double scn_DR_(ValidConvolution_updateOutput)(
@@ -71,11 +75,13 @@ extern "C" double scn_DR_(ValidConvolution_updateOutput)(
 
   SCN_INITIALIZE_AND_REFERENCE(Metadata<Dimension>, m)
   auto _rules = _m.getValidRuleBook(inputSize, filterSize, true);
-  uInt nActive = input_features->size[0];
+  uInt nActive = _m.getNActive(inputSize);
   THTensor_(resize2d)(output_features, nActive, weight->size[1]);
   if (not bias)
     THTensor_(zero)(output_features);
 
+  double flops = 0;
+  if (nActive) {
     auto iF = THTensor_(data)(input_features);
     auto oF = THTensor_(data)(output_features);
     auto ip = input_features->size[1];
@@ -85,10 +91,9 @@ extern "C" double scn_DR_(ValidConvolution_updateOutput)(
 
     Convolution_ForwardPass(iF, ip, ip, oF, op, op, w, b, _rules, nActive,
                             THBlas_(gemm));
-  double flops = 0;
     for (auto &r : _rules)
       flops += r.size() / 2 * ip * op;
-
+  }
   return flops;
 }
 
@@ -100,10 +105,11 @@ extern "C" void scn_DR_(ValidConvolution_backward)(
 
   SCN_INITIALIZE_AND_REFERENCE(Metadata<Dimension>, m)
   auto _rules = _m.getValidRuleBook(inputSize, filterSize, true);
-  uInt nActive = input_features->size[0];
+  uInt nActive = _m.getNActive(inputSize);
   THTensor_(resizeAs)(d_input_features, input_features);
   THTensor_(zero)(d_input_features);
 
+  if (nActive) {
     auto iF = THTensor_(data)(input_features);
     auto diF = THTensor_(data)(d_input_features);
     auto doF = THTensor_(data)(d_output_features);
@@ -115,6 +121,6 @@ extern "C" void scn_DR_(ValidConvolution_backward)(
 
     Convolution_BackwardPass(iF, diF, ip, ip, doF, op, op, w, dw, db, _rules,
                              nActive, THBlas_(gemm));
+  }
 }
-
 #endif

PyTorch/sparseconvnet/SCN/generic/CPU/SparseToDense.cpp

@@ -9,34 +9,32 @@
 #else
 #include "SparseToDense.h"
 
-extern "C" void scn_DR_(SparseToDense_updateOutput)(THLongTensor *inputSize,
-                                                    void **m,
-                                                    THTensor *input_features,
-                                                    THTensor *output_features,
-                                                    void *rulesBuffer) {
+extern "C" void scn_DR_(SparseToDense_updateOutput)(
+    THLongTensor *inputSize, void **m, THTensor *input_features,
+    THTensor *output_features, void *rulesBuffer, long nPlanes) {
+
+  SCN_INITIALIZE_AND_REFERENCE(Metadata<Dimension>, m)
 
-  SCN_INITIALIZE_AND_REFERENCE(Metadata<Dimension>, m) {
+  {
     long sz[Dimension + 2];
-    sz[0] = _m.inputSGs->size();
-    sz[1] = input_features->size[1];
-    for (int i = 0; i < Dimension; i++) {
-      auto x = THLongTensor_data(inputSize)[i];
-      sz[i + 2] = x;
-    }
+    sz[0] = _m.grids.begin()->second.size();
+    sz[1] = nPlanes; // input_features->size[1];
+    std::memcpy(sz + 2, THLongTensor_data(inputSize), sizeof(long) * Dimension);
     THTensor_(resizeNd)(output_features, Dimension + 2, sz, NULL);
     THTensor_(zero)(output_features);
   }
+  if (input_features->nDimension == 2) {
     auto _rules = _m.getSparseToDenseRuleBook(inputSize, true);
-  auto spatialVolume = _rules.size();
     uInt nPlanes = input_features->size[1];
     auto iF = THTensor_(data)(input_features);
     auto oF = THTensor_(data)(output_features);
-
+    long spatialVolume = THLongTensor_prodall(inputSize);
     for (auto &r : _rules) {
       uInt nHot = r.size() / 2;
       SparseToDense_ForwardPass<real>(iF, oF, nPlanes, spatialVolume, &r[0],
                                       nHot);
-    oF++;
+      oF += nPlanes * spatialVolume;
+    }
   }
 }
 extern "C" void scn_DR_(SparseToDense_updateGradInput)(
@@ -44,12 +42,12 @@ extern "C" void scn_DR_(SparseToDense_updateGradInput)(
     THTensor *d_input_features, THTensor *d_output_features,
     void *rulesBuffer) {
 
+  SCN_INITIALIZE_AND_REFERENCE(Metadata<Dimension>, m)
   THTensor_(resizeAs)(d_input_features, input_features);
   THTensor_(zero)(d_input_features);
-
-  SCN_INITIALIZE_AND_REFERENCE(Metadata<Dimension>, m)
   auto _rules = _m.getSparseToDenseRuleBook(inputSize, true);
-  auto spatialVolume = _rules.size();
+  if (input_features->nDimension == 2) {
+    long spatialVolume = THLongTensor_prodall(inputSize);
     uInt nPlanes = d_input_features->size[1];
     auto diF = THTensor_(data)(d_input_features);
     auto doF = THTensor_(data)(d_output_features);
@@ -58,7 +56,8 @@ extern "C" void scn_DR_(SparseToDense_updateGradInput)(
       uInt nHot = r.size() / 2;
       SparseToDense_BackwardPass<real>(diF, doF, nPlanes, spatialVolume, &r[0],
                                        nHot);
-    doF++;
+      doF += nPlanes * spatialVolume;
+    }
   }
 }
 #endif

PyTorch/sparseconvnet/SCN/generic/CPU/SparseToDense.h

@@ -10,27 +10,26 @@
 
 template <typename T>
 void SparseToDense_ForwardPass(T *input_features, T *output_features,
-                               uInt nPlanes, uInt spatialVolume,
-                               uInt* rules, int nHot) {
-
+                               uInt nPlanes, uInt spatialVolume, uInt *rules,
+                               int nHot) {
   for (uInt outSite = 0; outSite < nHot; outSite++) {
-                                   T *i = &input_features[rules[2 * outSite] * nPlanes];
-                                   uInt sample = rules[2 * outSite + 1];
+    T *i = input_features + rules[2 * outSite] * nPlanes;
+    T *o = output_features + rules[2 * outSite + 1];
     for (uInt plane = 0; plane < nPlanes; plane++)
-                                     output_features[(sample*nPlanes+plane)*spatialVolume]=i[plane];
+      o[plane * spatialVolume] = i[plane];
   }
 }
 
 template <typename T>
 void SparseToDense_BackwardPass(T *d_input_features, T *d_output_features,
-                                  uInt nPlanes, uInt spatialVolume,
-                                  uInt* rules, int nHot) {
+                                uInt nPlanes, uInt spatialVolume, uInt *rules,
+                                int nHot) {
 
   for (uInt outSite = 0; outSite < nHot; outSite++) {
-                                      T *di = &d_input_features[rules[2 * outSite] * nPlanes];
-                                      uInt sample = rules[2 * outSite + 1];
+    T *d_i = d_input_features + rules[2 * outSite] * nPlanes;
+    auto d_o = d_output_features + rules[2 * outSite + 1];
     for (uInt plane = 0; plane < nPlanes; plane++)
-                                        di[plane]=d_output_features[(sample*nPlanes+plane)*spatialVolume];
-                                    }
+      d_i[plane] = d_o[plane * spatialVolume];
   }
+}
 #endif /* CPU_SPARSETODENSE_H */

PyTorch/sparseconvnet/SCN/generic/GPU/AffineReluTrivialConvolution.cu

@@ -10,7 +10,6 @@
 #include "AffineReluTrivialConvolution.h"
 
 #include <algorithm>
-#include <iostream>
 
 extern "C" void scn_R_(AffineReluTrivialConvolution_updateOutput)(
     THCTensor *input_features, THCTensor *output_features,

PyTorch/sparseconvnet/SCN/generic/GPU/AffineReluTrivialConvolution.h

@@ -155,6 +155,27 @@ __global__ void dAffineReluTrivialConvolution_forwardB(
   }
 }
 
+#define FOO(T, K, V)                                                           \
+  {                                                                            \
+    if (input_nPlanes % K == 0 and output_nPlanes % K == 0) {                  \
+      uInt o = (nActive / K) * K;                                              \
+      if (o > 0)                                                               \
+        dAffineReluTrivialConvolution_forwardA<T, K, V> << <                   \
+            dim3(std::min(o / K, (uInt)512), output_nPlanes / K),              \
+            dim3(K, K / V), 0, THCState_getCurrentStream(state)>>>             \
+            (inFeatures, outFeatures, affineWeight, affineBias, convWeight,    \
+             input_nPlanes, input_stride, output_nPlanes, output_stride, o);   \
+      if (nActive > o)                                                         \
+        dAffineReluTrivialConvolution_forwardB<T, K, V> << <                   \
+            dim3(1, output_nPlanes / K), dim3(K, K / V), 0,                    \
+            THCState_getCurrentStream(state)>>>                                \
+            (inFeatures + o * input_stride, outFeatures + o * output_stride,   \
+             affineWeight, affineBias, convWeight, input_nPlanes,              \
+             input_stride, output_nPlanes, output_stride, nActive - o);        \
+      return;                                                                  \
+    }                                                                          \
+  }
+
 template <typename T>
 void dAffineReluTrivialConvolution_forward(T *inFeatures, T *outFeatures,
                                            T *affineWeight, T *affineBias,
@@ -162,92 +183,25 @@ void dAffineReluTrivialConvolution_forward(T *inFeatures, T *outFeatures,
                                            uInt input_stride,
                                            uInt output_nPlanes,
                                            uInt output_stride, uInt nActive) {
-  {
-    const uInt K = 64;
-    const uInt V = 16;
-    if (input_nPlanes % K == 0 and output_nPlanes % K == 0) {
-      uInt o = (nActive / K) * K;
-      if (o > 0)
-        dAffineReluTrivialConvolution_forwardA<
-            T, K, V><<<dim3(std::min(o / K, (uInt)512), output_nPlanes / K),
-                       dim3(K, K / V), 0, THCState_getCurrentStream(state)>>>(
-            inFeatures, outFeatures, affineWeight, affineBias, convWeight,
-            input_nPlanes, input_stride, output_nPlanes, output_stride, o);
-      if (nActive > o)
-        dAffineReluTrivialConvolution_forwardB<
-            T, K, V><<<dim3(1, output_nPlanes / K), dim3(K, K / V), 0,
-                       THCState_getCurrentStream(state)>>>(
-            inFeatures + o * input_stride, outFeatures + o * output_stride,
-            affineWeight, affineBias, convWeight, input_nPlanes, input_stride,
-            output_nPlanes, output_stride, nActive - o);
-      return;
-    }
-  }
-  {
-    const uInt K = 32;
-    const uInt V = 4;
-    if (input_nPlanes % K == 0 and output_nPlanes % K == 0) {
-      uInt o = (nActive / K) * K;
-      if (o > 0)
-        dAffineReluTrivialConvolution_forwardA<
-            T, K, V><<<dim3(std::min(o / K, (uInt)512), output_nPlanes / K),
-                       dim3(K, K / V), 0, THCState_getCurrentStream(state)>>>(
-            inFeatures, outFeatures, affineWeight, affineBias, convWeight,
-            input_nPlanes, input_stride, output_nPlanes, output_stride, o);
-      if (nActive > o)
-        dAffineReluTrivialConvolution_forwardB<
-            T, K, V><<<dim3(1, output_nPlanes / K), dim3(K, K / V), 0,
-                       THCState_getCurrentStream(state)>>>(
-            inFeatures + o * input_stride, outFeatures + o * output_stride,
-            affineWeight, affineBias, convWeight, input_nPlanes, input_stride,
-            output_nPlanes, output_stride, nActive - o);
-      return;
-    }
-  }
-  {
-    const uInt K = 16;
-    const uInt V = 4;
-    if (input_nPlanes % K == 0 and output_nPlanes % K == 0) {
-      uInt o = (nActive / K) * K;
-      if (o > 0)
-        dAffineReluTrivialConvolution_forwardA<
-            T, K, V><<<dim3(std::min(o / K, (uInt)512), output_nPlanes / K),
-                       dim3(K, K / V), 0, THCState_getCurrentStream(state)>>>(
-            inFeatures, outFeatures, affineWeight, affineBias, convWeight,
-            input_nPlanes, input_stride, output_nPlanes, output_stride, o);
-      if (nActive > o)
-        dAffineReluTrivialConvolution_forwardB<
-            T, K, V><<<dim3(1, output_nPlanes / K), dim3(K, K / V), 0,
-                       THCState_getCurrentStream(state)>>>(
-            inFeatures + o * input_stride, outFeatures + o * output_stride,
-            affineWeight, affineBias, convWeight, input_nPlanes, input_stride,
-            output_nPlanes, output_stride, nActive - o);
-      return;
-    }
-  }
-  {
-    const uInt K = 8;
-    const uInt V = 2;
-    if (input_nPlanes % K == 0 and output_nPlanes % K == 0) {
-      uInt o = (nActive / K) * K;
-      if (o > 0)
-        dAffineReluTrivialConvolution_forwardA<
-            T, K, V><<<dim3(std::min(o / K, (uInt)512), output_nPlanes / K),
-                       dim3(K, K / V), 0, THCState_getCurrentStream(state)>>>(
-            inFeatures, outFeatures, affineWeight, affineBias, convWeight,
-            input_nPlanes, input_stride, output_nPlanes, output_stride, o);
-      if (nActive > o)
-        dAffineReluTrivialConvolution_forwardB<
-            T, K, V><<<dim3(1, output_nPlanes / K), dim3(K, K / V), 0,
-                       THCState_getCurrentStream(state)>>>(
-            inFeatures + o * input_stride, outFeatures + o * output_stride,
-            affineWeight, affineBias, convWeight, input_nPlanes, input_stride,
-            output_nPlanes, output_stride, nActive - o);
-      return;
-    }
-  }
+
+  FOO(T, 64, 16)
+  FOO(T, 32, 8)
+  FOO(T, 16, 4)
+  FOO(T, 8, 2)
   assert(false);
 }
+template <>
+void dAffineReluTrivialConvolution_forward<double>(
+    double *inFeatures, double *outFeatures, double *affineWeight,
+    double *affineBias, double *convWeight, uInt input_nPlanes,
+    uInt input_stride, uInt output_nPlanes, uInt output_stride, uInt nActive) {
+
+  FOO(double, 32, 8)
+  FOO(double, 16, 4)
+  FOO(double, 8, 2)
+  assert(false);
+}
+#undef FOO
 
 // dOutput x W^T -> dInput and
 // Input^T x dOutput -> dW
@@ -449,84 +403,41 @@ __global__ void dAffineReluTrivialConvolution_backward_dW_B(
   atomicAdd(&dAffineBias[tx], dAB);
 }
 
+#define FOO(T, K, V)                                                           \
+  {                                                                            \
+    if (input_nPlanes % K == 0 and output_nPlanes % K == 0) {                  \
+      uInt o = (nActive / K) * K;                                              \
+      if (o > 0)                                                               \
+        dAffineReluTrivialConvolution_backward_dW_A<T, K, V> << <              \
+            dim3(std::min(o / K, (uInt)512), input_nPlanes / K),               \
+            dim3(K, K / V), 0, THCState_getCurrentStream(state)>>>             \
+            (inFeatures, dInFeatures, dOutFeatures, affineWeight,              \
+             dAffineWeight, affineBias, dAffineBias, convWeight, dConvWeight,  \
+             input_nPlanes, input_stride, output_nPlanes, output_stride, o,    \
+             additiveGrad);                                                    \
+      if (nActive > o)                                                         \
+        dAffineReluTrivialConvolution_backward_dW_B<T, K, V> << <              \
+            dim3(1, input_nPlanes / K), dim3(K, K / V), 0,                     \
+            THCState_getCurrentStream(state)>>>                                \
+            (inFeatures + o * input_stride, dInFeatures + o * input_stride,    \
+             dOutFeatures + o * output_stride, affineWeight, dAffineWeight,    \
+             affineBias, dAffineBias, convWeight, dConvWeight, input_nPlanes,  \
+             input_stride, output_nPlanes, output_stride, nActive - o,         \
+             additiveGrad);                                                    \
+      return;                                                                  \
+    }                                                                          \
+  }
+
 template <typename T>
 void dAffineReluTrivialConvolution_backward_dW(
     T *inFeatures, T *dInFeatures, T *dOutFeatures, T *affineWeight,
     T *dAffineWeight, T *affineBias, T *dAffineBias, T *convWeight,
     T *dConvWeight, uInt input_nPlanes, uInt input_stride, uInt output_nPlanes,
     uInt output_stride, uInt nActive, bool additiveGrad) {
-  {
-    const uInt K = 32;
-    const uInt V = 8;
-    if (input_nPlanes % K == 0 and output_nPlanes % K == 0) {
-      uInt o = (nActive / K) * K;
-      if (o > 0)
-        dAffineReluTrivialConvolution_backward_dW_A<
-            T, K, V><<<dim3(std::min(o / K, (uInt)512), input_nPlanes / K),
-                       dim3(K, K / V), 0, THCState_getCurrentStream(state)>>>(
-            inFeatures, dInFeatures, dOutFeatures, affineWeight, dAffineWeight,
-            affineBias, dAffineBias, convWeight, dConvWeight, input_nPlanes,
-            input_stride, output_nPlanes, output_stride, o, additiveGrad);
-      if (nActive > o)
-        dAffineReluTrivialConvolution_backward_dW_B<
-            T, K, V><<<dim3(1, input_nPlanes / K), dim3(K, K / V), 0,
-                       THCState_getCurrentStream(state)>>>(
-            inFeatures + o * input_stride, dInFeatures + o * input_stride,
-            dOutFeatures + o * output_stride, affineWeight, dAffineWeight,
-            affineBias, dAffineBias, convWeight, dConvWeight, input_nPlanes,
-            input_stride, output_nPlanes, output_stride, nActive - o,
-            additiveGrad);
-      return;
-    }
-  }
-  {
-    const uInt K = 16;
-    const uInt V = 4;
-    if (input_nPlanes % K == 0 and output_nPlanes % K == 0) {
-      uInt o = (nActive / K) * K;
-      if (o > 0)
-        dAffineReluTrivialConvolution_backward_dW_A<
-            T, K, V><<<dim3(std::min(o / K, (uInt)512), input_nPlanes / K),
-                       dim3(K, K / V), 0, THCState_getCurrentStream(state)>>>(
-            inFeatures, dInFeatures, dOutFeatures, affineWeight, dAffineWeight,
-            affineBias, dAffineBias, convWeight, dConvWeight, input_nPlanes,
-            input_stride, output_nPlanes, output_stride, o, additiveGrad);
-      if (nActive > o)
-        dAffineReluTrivialConvolution_backward_dW_B<
-            T, K, V><<<dim3(1, input_nPlanes / K), dim3(K, K / V), 0,
-                       THCState_getCurrentStream(state)>>>(
-            inFeatures + o * input_stride, dInFeatures + o * input_stride,
-            dOutFeatures + o * output_stride, affineWeight, dAffineWeight,
-            affineBias, dAffineBias, convWeight, dConvWeight, input_nPlanes,
-            input_stride, output_nPlanes, output_stride, nActive - o,
-            additiveGrad);
-      return;
-    }
-  }
-  {
-    const uInt K = 8;
-    const uInt V = 2;
-    if (input_nPlanes % K == 0 and output_nPlanes % K == 0) {
-      uInt o = (nActive / K) * K;
-      if (o > 0)
-        dAffineReluTrivialConvolution_backward_dW_A<
-            T, K, V><<<dim3(std::min(o / K, (uInt)512), input_nPlanes / K),
-                       dim3(K, K / V), 0, THCState_getCurrentStream(state)>>>(
-            inFeatures, dInFeatures, dOutFeatures, affineWeight, dAffineWeight,
-            affineBias, dAffineBias, convWeight, dConvWeight, input_nPlanes,
-            input_stride, output_nPlanes, output_stride, o, additiveGrad);
-      if (nActive > o)
-        dAffineReluTrivialConvolution_backward_dW_B<
-            T, K, V><<<dim3(1, input_nPlanes / K), dim3(K, K / V), 0,
-                       THCState_getCurrentStream(state)>>>(
-            inFeatures + o * input_stride, dInFeatures + o * input_stride,
-            dOutFeatures + o * output_stride, affineWeight, dAffineWeight,
-            affineBias, dAffineBias, convWeight, dConvWeight, input_nPlanes,
-            input_stride, output_nPlanes, output_stride, nActive - o,
-            additiveGrad);
-      return;
-    }
-  }
+  FOO(T, 32, 8)
+  FOO(T, 16, 4)
+  FOO(T, 8, 2)
 }
+#undef FOO
 
 #endif

PyTorch/sparseconvnet/SCN/generic/GPU/BatchNormalization.cu

@@ -30,13 +30,14 @@ extern "C" void scn_R_(BatchNormalization_updateOutput)(
     real leakiness) {
 
   THCTensor_(resizeAs)(state, output_features, input_features);
+  if (input_features->nDimension == 2) {
     auto nActive = input_features->size[0];
     auto nPlanes = input_features->size[1];
     auto input_stride = input_features->stride[0];
     auto output_stride = output_features->stride[0];
-
     BN_F_MACRO(16)
     else BN_F_MACRO(12) else BN_F_MACRO(8) else BN_F_MACRO(4) else BN_F_MACRO(1)
+  }
 }
 
 extern "C" void scn_R_(BatchNormalizationInTensor_updateOutput)(
@@ -44,14 +45,14 @@ extern "C" void scn_R_(BatchNormalizationInTensor_updateOutput)(
     THCTensor *saveInvStd, THCTensor *runningMean, THCTensor *runningVar,
     THCTensor *weight, THCTensor *bias, real eps, real momentum, bool train,
     real leakiness) {
-
+  if (input_features->nDimension == 2) {
     auto nActive = input_features->size[0];
     auto nPlanes = input_features->size[1];
     auto input_stride = input_features->stride[0];
     auto output_stride = output_features->stride[0];
-
     BN_F_MACRO(16)
     else BN_F_MACRO(12) else BN_F_MACRO(8) else BN_F_MACRO(4) else BN_F_MACRO(1)
+  }
 }
 
 #undef BN_F_MACRO
@@ -81,12 +82,13 @@ extern "C" void scn_R_(BatchNormalization_backward)(
     THCTensor *d_weight, THCTensor *d_bias, real leakiness) {
 
   THCTensor_(resizeAs)(state, d_input_features, d_output_features);
+  if (input_features->nDimension == 2) {
     auto nActive = input_features->size[0];
     auto nPlanes = input_features->size[1];
     auto input_stride = input_features->stride[0];
     auto output_stride = output_features->stride[0];
-
     BN_B_MACRO(16)
     else BN_B_MACRO(12) else BN_B_MACRO(8) else BN_B_MACRO(4) else BN_B_MACRO(1)
+  }
 }
 #endif

PyTorch/sparseconvnet/SCN/generic/GPU/Convolution.cu

@@ -25,12 +25,13 @@ extern "C" double scn_DR_(Convolution_updateOutput)(
   if (not bias)
     THCTensor_(zero)(state, output_features);
 
+  double flops = 0;
+  if (nActive) {
     auto iF = THCTensor_(data)(state, input_features);
     auto oF = THCTensor_(data)(state, output_features);
     auto ip = input_features->size[1];
     auto op = output_features->size[1];
     auto w = THCTensor_(data)(state, weight);
-  double flops = 0;
 
     if (bias) {
       auto b = THCTensor_(data)(state, bias);
@@ -47,6 +48,7 @@ extern "C" double scn_DR_(Convolution_updateOutput)(
         dConvolution_forward2<real>(iF, oF, w, rbB, nHotB, ip, ip, op, op,
                                     THCState_getCurrentStream(state));
         , w += c; flops += nHotB * c;)
+  }
   return flops;
 }
 
@@ -63,6 +65,7 @@ extern "C" void scn_DR_(Convolution_backward)(
   THCTensor_(resizeAs)(state, d_input_features, input_features);
   THCTensor_(zero)(state, d_input_features);
 
+  if (nActive) {
     auto iF = THCTensor_(data)(state, input_features);
     auto diF = THCTensor_(data)(state, d_input_features);
     auto doF = THCTensor_(data)(state, d_output_features);
@@ -71,9 +74,9 @@ extern "C" void scn_DR_(Convolution_backward)(
     auto w = THCTensor_(data)(state, weight);
     auto dw = THCTensor_(data)(state, d_weight);
     uInt c = ip * op;
-  RULEBOOKITERATOR(
-      dConvolution_backward_dW2<real>(iF, diF, doF, w, dw, rbB, nHotB, ip, ip,
-                                      op, op, THCState_getCurrentStream(state));
+    RULEBOOKITERATOR(dConvolution_backward_dW2<real>(
+                         iF, diF, doF, w, dw, rbB, nHotB, ip, ip, op, op,
+                         THCState_getCurrentStream(state));
                      , w += c; dw += c;)
 
     if (d_bias) {
@@ -81,6 +84,7 @@ extern "C" void scn_DR_(Convolution_backward)(
       Convolution_bp_bias(doF, db, op, op, nActive,
                           THCState_getCurrentStream(state));
     }
+  }
 }
 
 extern "C" double scn_DR_(ValidConvolution_updateOutput)(
@@ -89,17 +93,18 @@ extern "C" double scn_DR_(ValidConvolution_updateOutput)(
     THCTensor *bias, long filterVolume, THCITensor *rulesBuffer) {
   SCN_INITIALIZE_AND_REFERENCE(Metadata<Dimension>, m)
   auto _rules = _m.getValidRuleBook(inputSize, filterSize, true);
-  uInt nActive = input_features->size[0];
+  uInt nActive = _m.getNActive(inputSize);
   THCTensor_(resize2d)(state, output_features, nActive, weight->size[1]);
   if (not bias)
     THCTensor_(zero)(state, output_features);
 
+  double flops = 0;
+  if (nActive) {
     auto iF = THCTensor_(data)(state, input_features);
     auto oF = THCTensor_(data)(state, output_features);
     auto ip = input_features->size[1];
     auto op = output_features->size[1];
     auto w = THCTensor_(data)(state, weight);
-  double flops = 0;
 
     if (bias) {
       auto b = THCTensor_(data)(state, bias);
@@ -116,6 +121,7 @@ extern "C" double scn_DR_(ValidConvolution_updateOutput)(
         dConvolution_forward2<real>(iF, oF, w, rbB, nHotB, ip, ip, op, op,
                                     THCState_getCurrentStream(state));
         , w += c; flops += nHotB * c;)
+  }
   return flops;
 }
 
@@ -126,10 +132,11 @@ extern "C" void scn_DR_(ValidConvolution_backward)(
     THCTensor *d_bias, long filterVolume, THCITensor *rulesBuffer) {
   SCN_INITIALIZE_AND_REFERENCE(Metadata<Dimension>, m)
   auto _rules = _m.getValidRuleBook(inputSize, filterSize, true);
-  uInt nActive = input_features->size[0];
+  uInt nActive = _m.getNActive(inputSize);
   THCTensor_(resizeAs)(state, d_input_features, input_features);
   THCTensor_(zero)(state, d_input_features);
 
+  if (nActive) {
     auto iF = THCTensor_(data)(state, input_features);
     auto diF = THCTensor_(data)(state, d_input_features);
     auto doF = THCTensor_(data)(state, d_output_features);
@@ -138,9 +145,9 @@ extern "C" void scn_DR_(ValidConvolution_backward)(
     auto w = THCTensor_(data)(state, weight);
     auto dw = THCTensor_(data)(state, d_weight);
     uInt c = ip * op;
-  RULEBOOKITERATOR(
-      dConvolution_backward_dW2<real>(iF, diF, doF, w, dw, rbB, nHotB, ip, ip,
-                                      op, op, THCState_getCurrentStream(state));
+    RULEBOOKITERATOR(dConvolution_backward_dW2<real>(
+                         iF, diF, doF, w, dw, rbB, nHotB, ip, ip, op, op,
+                         THCState_getCurrentStream(state));
                      , w += c; dw += c;)
 
     if (d_bias) {
@@ -148,6 +155,7 @@ extern "C" void scn_DR_(ValidConvolution_backward)(
       Convolution_bp_bias(doF, db, op, op, nActive,
                           THCState_getCurrentStream(state));
     }
+  }
 }
 
 #endif

PyTorch/sparseconvnet/SCN/generic/GPU/Convolution.h

@@ -184,7 +184,7 @@ dConvolution_KMxKN_forwardB(T *inFeatures, T *outFeatures, T *w, uInt *rules,
   }
 }
 
-#define FOO(K, V)                                                              \
+#define FOO(T, K, V)                                                           \
   {                                                                            \
     if (input_nPlanes % K == 0 and output_nPlanes % K == 0) {                  \
       uInt o = (nHot / K) * K;                                                 \
@@ -208,10 +208,21 @@ void dConvolution_forward(T *inFeatures, T *outFeatures, T *w, uInt *rules,
                           uInt nHot, uInt input_nPlanes, uInt input_stride,
                           uInt output_nPlanes, uInt output_stride,
                           cudaStream_t stream) {
-  FOO(64, 16)
-  FOO(32, 8)
-  FOO(16, 4)
-  FOO(8, 2)
+  FOO(T, 64, 16)
+  FOO(T, 32, 8)
+  FOO(T, 16, 4)
+  FOO(T, 8, 2)
+  assert(false);
+}
+template <>
+void dConvolution_forward<double>(double *inFeatures, double *outFeatures,
+                                  double *w, uInt *rules, uInt nHot,
+                                  uInt input_nPlanes, uInt input_stride,
+                                  uInt output_nPlanes, uInt output_stride,
+                                  cudaStream_t stream) {
+  FOO(double, 32, 8)
+  FOO(double, 16, 4)
+  FOO(double, 8, 2)
   assert(false);
 }
 #undef FOO
@@ -378,7 +389,7 @@ dConvolution_KMxKN_backward_dW_B(T *inFeatures, T *dInFeatures, T *dOutFeatures,
   }
 }
 
-#define FOO(K, V)                                                              \
+#define FOO(T, K, V)                                                           \
   {                                                                            \
     if (input_nPlanes % K == 0 and output_nPlanes % K == 0) {                  \
       uInt o = (nHot / K) * K;                                                 \
@@ -404,9 +415,9 @@ void dConvolution_backward_dW(T *inFeatures, T *dInFeatures, T *dOutFeatures,
                               uInt input_nPlanes, uInt input_stride,
                               uInt output_nPlanes, uInt output_stride,
                               cudaStream_t stream) {
-  FOO(32, 8)
-  FOO(16, 4)
-  FOO(8, 2)
+  FOO(T, 32, 8)
+  FOO(T, 16, 4)
+  FOO(T, 8, 2)
   assert(false);
 }
 #undef FOO

PyTorch/sparseconvnet/SCN/generic/GPU/Deconvolution.h

@@ -153,7 +153,7 @@ dDeconvolution_KMxKN_forwardB(T *inFeatures, T *outFeatures, T *w, uInt *rules,
   }
 }
 
-#define FOO(K, V)                                                              \
+#define FOO(T, K, V)                                                           \
   {                                                                            \
     if (input_nPlanes % K == 0 and output_nPlanes % K == 0) {                  \
       uInt o = (nHot / K) * K;                                                 \
@@ -177,10 +177,21 @@ void dDeconvolution_forward(T *inFeatures, T *outFeatures, T *w, uInt *rules,
                             uInt nHot, uInt input_nPlanes, uInt input_stride,
                             uInt output_nPlanes, uInt output_stride,
                             cudaStream_t stream) {
-  FOO(64, 16)
-  FOO(32, 8)
-  FOO(16, 4)
-  FOO(8, 2)
+  FOO(T, 64, 16)
+  FOO(T, 32, 8)
+  FOO(T, 16, 4)
+  FOO(T, 8, 2)
+  assert(false);
+}
+template <>
+void dDeconvolution_forward<double>(double *inFeatures, double *outFeatures,
+                                    double *w, uInt *rules, uInt nHot,
+                                    uInt input_nPlanes, uInt input_stride,
+                                    uInt output_nPlanes, uInt output_stride,
+                                    cudaStream_t stream) {
+  FOO(double, 32, 8)
+  FOO(double, 16, 4)
+  FOO(double, 8, 2)
   assert(false);
 }
 #undef FOO
@@ -345,7 +356,7 @@ __global__ void dDeconvolution_KMxKN_backward_dW_B(
   }
 }
 
-#define FOO(K, V)                                                              \
+#define FOO(T, K, V)                                                           \
   {                                                                            \
     if (input_nPlanes % K == 0 and output_nPlanes % K == 0) {                  \
       uInt o = (nHot / K) * K;                                                 \
@@ -371,9 +382,9 @@ void dDeconvolution_backward_dW(T *inFeatures, T *dInFeatures, T *dOutFeatures,
                                 uInt input_nPlanes, uInt input_stride,
                                 uInt output_nPlanes, uInt output_stride,
                                 cudaStream_t stream) {
-  FOO(32, 8)
-  FOO(16, 4)
-  FOO(8, 2)
+  FOO(T, 32, 8)
+  FOO(T, 16, 4)
+  FOO(T, 8, 2)
   assert(false);
 }
 #undef FOO

PyTorch/sparseconvnet/SCN/generic/GPU/SparseToDense.cu

@@ -9,50 +9,54 @@
 #else
 #include "SparseToDense.h"
 
-extern "C" void scn_DR_(SparseToDense_updateOutput)(THLongTensor *inputSize,
-                                                    void **m,
-                                                    THCTensor *input_features,
-                                                    THCTensor *output_features,
-                                                    THCITensor *rulesBuffer) {
+extern "C" void scn_DR_(SparseToDense_updateOutput)(
+    THLongTensor *inputSize, void **m, THCTensor *input_features,
+    THCTensor *output_features, THCITensor *rulesBuffer, long nPlanes) {
 
-  SCN_INITIALIZE_AND_REFERENCE(Metadata<Dimension>, m) {
+  SCN_INITIALIZE_AND_REFERENCE(Metadata<Dimension>, m)
+  long spatialVolume = 1;
+  {
     long sz[Dimension + 2];
-    sz[0] = _m.inputSGs->size();
-    sz[1] = input_features->size[1];
+    sz[0] = _m.grids.begin()->second.size();
+    sz[1] = nPlanes; // input_features->size[1];
     for (int i = 0; i < Dimension; i++) {
       auto x = THLongTensor_data(inputSize)[i];
       sz[i + 2] = x;
+      spatialVolume *= x;
     }
     THCTensor_(resizeNd)(state, output_features, Dimension + 2, sz, NULL);
     THCTensor_(zero)(state, output_features);
   }
+  if (input_features->nDimension == 2) {
     auto _rules = _m.getSparseToDenseRuleBook(inputSize, true);
-  auto spatialVolume = _rules.size();
     uInt nPlanes = input_features->size[1];
     auto iF = THCTensor_(data)(state, input_features);
     auto oF = THCTensor_(data)(state, output_features);
     RULEBOOKITERATOR(
-      SparseToDense_ForwardPass<real>(THCState_getCurrentStream(state), iF, oF,
-                                      nPlanes, spatialVolume, rbB, nHotB);
-      , oF++;) // todo check ++ or +=spatialVolume????zzz
+        SparseToDense_ForwardPass<real>(THCState_getCurrentStream(state), iF,
+                                        oF, nPlanes, spatialVolume, rbB, nHotB);
+        , oF += nPlanes * spatialVolume;)
+  }
 }
 extern "C" void scn_DR_(SparseToDense_updateGradInput)(
     THLongTensor *inputSize, void **m, THCTensor *input_features,
     THCTensor *d_input_features, THCTensor *d_output_features,
     THCITensor *rulesBuffer) {
 
+  SCN_INITIALIZE_AND_REFERENCE(Metadata<Dimension>, m)
   THCTensor_(resizeAs)(state, d_input_features, input_features);
   THCTensor_(zero)(state, d_input_features);
 
-  SCN_INITIALIZE_AND_REFERENCE(Metadata<Dimension>, m)
+  if (input_features->nDimension == 2) {
     auto _rules = _m.getSparseToDenseRuleBook(inputSize, true);
-  auto spatialVolume = _rules.size();
+    long spatialVolume = THLongTensor_prodall(inputSize);
     uInt nPlanes = d_input_features->size[1];
     auto diF = THCTensor_(data)(state, d_input_features);
     auto doF = THCTensor_(data)(state, d_output_features);
-  RULEBOOKITERATOR(
-      SparseToDense_BackwardPass<real>(THCState_getCurrentStream(state), diF,
-                                       doF, nPlanes, spatialVolume, rbB, nHotB);
-      , doF++;)
+    RULEBOOKITERATOR(SparseToDense_BackwardPass<real>(
+                         THCState_getCurrentStream(state), diF, doF, nPlanes,
+                         spatialVolume, rbB, nHotB);
+                     , doF += nPlanes * spatialVolume;)
+  }
 }
 #endif

PyTorch/sparseconvnet/SCN/generic/GPU/SparseToDense.h

@@ -12,7 +12,8 @@
 // NTX must be >=2 so r is filled properly
 template <typename T, uInt NTX, uInt NTY>
 __global__ void SparseToDense_fp(T *input_features, T *output_features,
-                                uInt nPlanes, uInt spatialVolume, uInt *rules, uInt nHot) {
+                                 uInt nPlanes, uInt spatialVolume, uInt *rules,
+                                 uInt nHot) {
   __shared__ uInt r[NTY * 2];
   for (uInt n = blockIdx.x * NTY; n < nHot; n += gridDim.x * NTY) {
     {
@@ -22,10 +23,10 @@ __global__ void SparseToDense_fp(T *input_features, T *output_features,
     }
     __syncthreads();
     if (n + threadIdx.y < nHot) {
-      T *i = &input_features[r[2 * threadIdx.y] * nPlanes];
-      T *o = &output_features[r[2*threadIdx.y+1]*spatialVolume*nPlanes];
+      T *i = input_features + r[2 * threadIdx.y] * nPlanes;
+      T *o = output_features + r[2 * threadIdx.y + 1];
       for (uInt plane = threadIdx.x; plane < nPlanes; plane += NTX)
-      o[plane*spatialVolume]=i[plane];
+        o[plane * spatialVolume] = i[plane];
     }
     __syncthreads();
   }
@@ -34,15 +35,15 @@ __global__ void SparseToDense_fp(T *input_features, T *output_features,
 template <typename T>
 void SparseToDense_ForwardPass(cudaStream_t stream, T *input_features,
                                T *output_features, uInt nPlanes,
-                              uInt spatialVolume,
-                              uInt *rules, uInt nHot) {
-  SparseToDense_fp<T, 32, 32><<<32, dim3(32, 32), 0, stream>>>(
-      input_features, output_features, nPlanes, spatialVolume,  rules, nHot);
+                               uInt spatialVolume, uInt *rules, uInt nHot) {
+  SparseToDense_fp<T, 32, 32> << <32, dim3(32, 32), 0, stream>>>
+      (input_features, output_features, nPlanes, spatialVolume, rules, nHot);
 }
 // NTX must be >=2 so r is filled properly
 template <typename T, uInt NTX, uInt NTY>
 __global__ void SparseToDense_bp(T *d_input_features, T *d_output_features,
-                                uInt nPlanes, uInt spatialVolume, uInt *rules, uInt nHot) {
+                                 uInt nPlanes, uInt spatialVolume, uInt *rules,
+                                 uInt nHot) {
   __shared__ uInt r[NTY * 2];
   for (uInt n = blockIdx.x * NTY; n < nHot; n += gridDim.x * NTY) {
     {
@@ -52,10 +53,10 @@ __global__ void SparseToDense_bp(T *d_input_features, T *d_output_features,
     }
     __syncthreads();
     if (n + threadIdx.y < nHot) {
-      T *i = &d_input_features[r[2 * threadIdx.y] * nPlanes];
-      T *o = &d_output_features[r[2*threadIdx.y+1]*spatialVolume*nPlanes];
+      T *d_i = d_input_features + r[2 * threadIdx.y] * nPlanes;
+      T *d_o = d_output_features + r[2 * threadIdx.y + 1];
       for (uInt plane = threadIdx.x; plane < nPlanes; plane += NTX)
-      i[plane]=o[plane*spatialVolume];
+        d_i[plane] = d_o[plane * spatialVolume];
     }
     __syncthreads();
   }
@@ -64,9 +65,9 @@ __global__ void SparseToDense_bp(T *d_input_features, T *d_output_features,
 template <typename T>
 void SparseToDense_BackwardPass(cudaStream_t stream, T *d_input_features,
                                 T *d_output_features, uInt nPlanes,
-                              uInt spatialVolume,
-                              uInt *rules, uInt nHot) {
-  SparseToDense_bp<T, 32, 32><<<32, dim3(32, 32), 0, stream>>>(
-      d_input_features, d_output_features, nPlanes, spatialVolume,  rules, nHot);
+                                uInt spatialVolume, uInt *rules, uInt nHot) {
+  SparseToDense_bp<T, 32, 32> << <32, dim3(32, 32), 0, stream>>>
+      (d_input_features, d_output_features, nPlanes, spatialVolume, rules,
+       nHot);
 }
 #endif /* GPU_SPARSETODENSE_H */

PyTorch/sparseconvnet/SCN/generic/GPU/THGenerateCudaFloatTypes.h

@@ -27,4 +27,22 @@
 #undef TH_REAL_IS_FLOAT
 #undef THBLAS_GEMM
 
+// double
+// #define real double
+// #define accreal double
+// #define Real Double
+// #define CReal CudaDouble
+// #define TH_REAL_IS_DOUBLE
+// #define THBLAS_GEMM THCudaBlas_Dgemm
+
+// #line 1 TH_GENERIC_FILE
+// #include TH_GENERIC_FILE
+
+// #undef accreal
+// #undef real
+// #undef Real
+// #undef CReal
+// #undef TH_REAL_IS_DOUBLE
+// #undef THBLAS_GEMM
+
 #undef TH_GENERIC_FILE

PyTorch/sparseconvnet/SCN/generic/Geometry/ConvolutionRules.h

@@ -103,20 +103,26 @@ uInt Convolution_InputSgsToRulesAndOutputSgs_OMP(
   return output_nActive;
 }
 
-// for each site in filterVolume, list of (inputFeatureNumber,batchIdx) pairs
+// for each active site, list of (inputFeatureNumber,batchIdx, spatialOffset)
+// triples
 template <uInt dimension>
 void SparseToDense_InputSgsToRulesAndOutputSgs(
     SparseGrids<dimension> &input_SGs, RuleBook &rules, long *spatialSize) {
   uInt batchSize = input_SGs.size();
-  SparseGrids<dimension> output_SGs(batchSize);
-  std::vector<long> ones(dimension, 1);
   rules.clear();
-  for (uInt i = 0; i < batchSize; i++) {
-    auto &iSG = input_SGs[i];
-    auto &oSG = output_SGs[i];
-    oSG.ctr = i; // batchIdx
-    Convolution_InputSgToRulesAndOutputSg<dimension>(
-        iSG, oSG, rules, spatialSize, &ones[0], spatialSize, &ones[0]);
+  rules.resize(batchSize);
+  Point<dimension> lb, ub;
+  for (int i = 0; i < dimension; ++i) {
+    lb[i] = 0;
+    ub[i] = spatialSize[i] - 1;
+  }
+  auto region = RectangularRegion<dimension>(lb, ub);
+  for (uInt batchIdx = 0; batchIdx < batchSize; batchIdx++) {
+    auto &iSG = input_SGs[batchIdx];
+    for (auto const &inIter : iSG.mp) {
+      rules[batchIdx].push_back(inIter.second + iSG.ctr);
+      rules[batchIdx].push_back(region.offset(inIter.first));
+    }
   }
 }
 
@@ -124,33 +130,21 @@ template <uInt dimension>
 void SparseToDense_InputSgsToRulesAndOutputSgs_OMP(
     SparseGrids<dimension> &input_SGs, RuleBook &rules, long *spatialSize) {
   uInt batchSize = input_SGs.size();
-  SparseGrids<dimension> output_SGs(batchSize);
-  std::vector<long> ones(dimension, 1);
   rules.clear();
-  rules.resize(volume<dimension>(spatialSize));
-  std::vector<RuleBook> rbs(batchSize);
-  {
-    uInt i;
-#pragma omp parallel for private(i)
-    for (i = 0; i < batchSize; i++) {
-      output_SGs[i].ctr = i; // batchIdx
-      Convolution_InputSgToRulesAndOutputSg<dimension>(
-          input_SGs[i], output_SGs[i], rbs[i], spatialSize, &ones[0],
-          spatialSize, &ones[0]);
-    }
-  }
-  {
-    uInt i;
-#pragma omp parallel for private(i)
-    for (i = 0; i < rules.size(); i++) {
-      auto &R = rules[i];
-      for (uInt j = 0; j < batchSize; j++) {
-        auto &r = rbs[j][i];
-        for (uInt k = 0; k < r.size();) {
-          R.push_back(r[k++]);
-          R.push_back(r[k++]);
-        }
-      }
+  rules.resize(batchSize);
+  Point<dimension> lb, ub;
+  for (int i = 0; i < dimension; ++i) {
+    lb[i] = 0;
+    ub[i] = spatialSize[i] - 1;
+  }
+  auto region = RectangularRegion<dimension>(lb, ub);
+  uInt batchIdx;
+#pragma omp parallel for private(batchIdx)
+  for (batchIdx = 0; batchIdx < batchSize; batchIdx++) {
+    auto &iSG = input_SGs[batchIdx];
+    for (auto const &inIter : iSG.mp) {
+      rules[batchIdx].push_back(inIter.second + iSG.ctr);
+      rules[batchIdx].push_back(region.offset(inIter.first));
     }
   }
 }

PyTorch/sparseconvnet/SCN/generic/Geometry/Metadata.cpp

@@ -125,16 +125,15 @@ extern "C" void scn_D_(getSpatialLocations)(void **m, THLongTensor *spatialSize,
 }
 extern "C" void
 scn_D_(createMetadataForDenseToSparse)(void **m, THLongTensor *spatialSize_,
-                                       THLongTensor *pad_, THLongTensor *nz_,
-                                       long batchSize) {
+                                       THLongTensor *nz_, long batchSize) {
   SCN_INITIALIZE_AND_REFERENCE(Metadata<Dimension>, m)
+  _m.clear();
   _m.setInputSpatialSize(spatialSize_);
   _m.inputSGs->resize(batchSize);
   auto &nActive = *_m.inputNActive;
   nActive = nz_->size[0];
 
   auto nz = THLongTensor_data(nz_);
-  auto pad = THLongTensor_data(pad_);
   auto spatialSize = THLongTensor_data(spatialSize_);
 
   std::vector<uInt> br(batchSize + 1);
@@ -157,8 +156,7 @@ scn_D_(createMetadataForDenseToSparse)(void **m, THLongTensor *spatialSize_,
     for (uInt i = br[b]; i < br[b + 1]; i++) {
       Point<Dimension> x;
       for (uInt j = 0; j < Dimension; j++) {
-        x[j] = nz[i * (Dimension + 1) + j + 1] +
-               pad[b * Dimension + j]; // 0-indexed
+        x[j] = nz[i * (Dimension + 1) + j + 1]; // 0-indexed
       }
       sg.mp[x] = i;
     }
@@ -281,6 +279,7 @@ extern "C" void scn_D_(generateRuleBooks2s2)(void **m) {
       p2[i] = p3[i] = inS[i] = outS[i];
   }
 }
+
 extern "C" void scn_D_(freeMetadata)(void **m) {
   SCN_DELETE(Metadata<Dimension>, m)
 }

PyTorch/sparseconvnet/SCN/generic/Geometry/Metadata.h

@@ -11,7 +11,6 @@
 #include "ActivePoolingRules.h"
 #include "ConvolutionRules.h"
 #include "ValidConvolutionRules.h"
-#include <iostream>
 #include <tuple>
 #include <unordered_map>
 
@@ -40,6 +39,18 @@ public:
   uInt *inputNActive;
 
   Metadata() {}
+  void clear() {
+    nActive.clear();
+    grids.clear();
+    activePoolingRuleBooks.clear();
+    validRuleBooks.clear();
+    ruleBooks.clear();
+    sparseToDenseRuleBooks.clear();
+    inputSGs = nullptr;
+    inputSG = nullptr;
+    inputNActive = nullptr;
+  }
+
   void setInputSpatialSize(THLongTensor *spatialSize) {
     inputSpatialSize = LongTensorToPoint<dimension>(spatialSize);
     inputSGs = &grids[inputSpatialSize];

PyTorch/sparseconvnet/SCN/generic/Geometry/ValidConvolutionRules.h

@@ -6,8 +6,6 @@
 
 #ifndef VALIDCONVOLUTIONRULES_H
 #define VALIDCONVOLUTIONRULES_H
-#include<iostream>
-
 
 // Full input region for an output point
 template <uInt dimension>
@@ -26,8 +24,8 @@ InputRegionCalculator_Valid(const Point<dimension> &output, long *size) {
 // rules is used to carry out the "lowering" whilst carrying out the convolution
 
 template <uInt dimension>
-double ValidConvolution_SgToRules(SparseGrid<dimension> &grid,
-                                    RuleBook &rules, long *size) {
+double ValidConvolution_SgToRules(SparseGrid<dimension> &grid, RuleBook &rules,
+                                  long *size) {
   uInt sd = volume<dimension>(size);
   double countActiveInputs = 0;
   for (auto const &outputIter : grid.mp) {
@@ -48,8 +46,8 @@ double ValidConvolution_SgToRules(SparseGrid<dimension> &grid,
 }
 
 template <uInt dimension>
-uInt ValidConvolution_SgsToRules(SparseGrids<dimension> &SGs,
-                                   RuleBook &rules, long *size) {
+uInt ValidConvolution_SgsToRules(SparseGrids<dimension> &SGs, RuleBook &rules,
+                                 long *size) {
   uInt sd = volume<dimension>(size);
   uInt countActiveInputs = 0;
   rules.clear();

PyTorch/sparseconvnet/SCN/header_cpu.h

@@ -8,13 +8,12 @@ long scn_readPtr(void **ptr);
 void scn_writePtr(long p, void **ptr);
 double scn_ruleBookBits(void);
 void scn_2_drawCurve(void **m, THFloatTensor *features, THFloatTensor *stroke);
-  double scn_1_addSampleFromThresholdedTensor(
+double scn_1_addSampleFromThresholdedTensor(
   void **m, THFloatTensor *features_, THFloatTensor *tensor_,
   THLongTensor *offset_, THLongTensor *spatialSize_, float threshold);
 void scn_1_batchAddSample(void **m);
 void scn_1_createMetadataForDenseToSparse(
-  void **m, THLongTensor *spatialSize_, THLongTensor *pad, THLongTensor *nz,
-  long batchSize);
+  void **m, THLongTensor *spatialSize_, THLongTensor *nz, long batchSize);
 void scn_1_freeMetadata(void **metadata);
 void scn_1_generateRuleBooks3s2(void **m);
 void scn_1_generateRuleBooks2s2(void **m);
@@ -25,13 +24,13 @@ void scn_1_setInputSpatialLocations(void **m, THFloatTensor *features,
   THLongTensor *locations, THFloatTensor *vecs, _Bool overwrite);
 void scn_1_getSpatialLocations(void **m, THLongTensor *spatialSize,
     THLongTensor *locations);
-  double scn_2_addSampleFromThresholdedTensor(
+
+double scn_2_addSampleFromThresholdedTensor(
   void **m, THFloatTensor *features_, THFloatTensor *tensor_,
   THLongTensor *offset_, THLongTensor *spatialSize_, float threshold);
 void scn_2_batchAddSample(void **m);
 void scn_2_createMetadataForDenseToSparse(
-  void **m, THLongTensor *spatialSize_, THLongTensor *pad, THLongTensor *nz,
-  long batchSize);
+  void **m, THLongTensor *spatialSize_, THLongTensor *nz, long batchSize);
 void scn_2_freeMetadata(void **metadata);
 void scn_2_generateRuleBooks3s2(void **m);
 void scn_2_generateRuleBooks2s2(void **m);
@@ -42,13 +41,13 @@ void scn_2_setInputSpatialLocations(void **m, THFloatTensor *features,
   THLongTensor *locations, THFloatTensor *vecs, _Bool overwrite);
 void scn_2_getSpatialLocations(void **m, THLongTensor *spatialSize,
     THLongTensor *locations);
-  double scn_3_addSampleFromThresholdedTensor(
+
+double scn_3_addSampleFromThresholdedTensor(
   void **m, THFloatTensor *features_, THFloatTensor *tensor_,
   THLongTensor *offset_, THLongTensor *spatialSize_, float threshold);
 void scn_3_batchAddSample(void **m);
 void scn_3_createMetadataForDenseToSparse(
-  void **m, THLongTensor *spatialSize_, THLongTensor *pad, THLongTensor *nz,
-  long batchSize);
+  void **m, THLongTensor *spatialSize_, THLongTensor *nz, long batchSize);
 void scn_3_freeMetadata(void **metadata);
 void scn_3_generateRuleBooks3s2(void **m);
 void scn_3_generateRuleBooks2s2(void **m);
@@ -59,13 +58,13 @@ void scn_3_setInputSpatialLocations(void **m, THFloatTensor *features,
   THLongTensor *locations, THFloatTensor *vecs, _Bool overwrite);
 void scn_3_getSpatialLocations(void **m, THLongTensor *spatialSize,
     THLongTensor *locations);
-  double scn_4_addSampleFromThresholdedTensor(
+
+double scn_4_addSampleFromThresholdedTensor(
   void **m, THFloatTensor *features_, THFloatTensor *tensor_,
   THLongTensor *offset_, THLongTensor *spatialSize_, float threshold);
 void scn_4_batchAddSample(void **m);
 void scn_4_createMetadataForDenseToSparse(
-  void **m, THLongTensor *spatialSize_, THLongTensor *pad, THLongTensor *nz,
-  long batchSize);
+  void **m, THLongTensor *spatialSize_, THLongTensor *nz, long batchSize);
 void scn_4_freeMetadata(void **metadata);
 void scn_4_generateRuleBooks3s2(void **m);
 void scn_4_generateRuleBooks2s2(void **m);
@@ -76,13 +75,13 @@ void scn_4_setInputSpatialLocations(void **m, THFloatTensor *features,
   THLongTensor *locations, THFloatTensor *vecs, _Bool overwrite);
 void scn_4_getSpatialLocations(void **m, THLongTensor *spatialSize,
     THLongTensor *locations);
-  double scn_5_addSampleFromThresholdedTensor(
+
+double scn_5_addSampleFromThresholdedTensor(
   void **m, THFloatTensor *features_, THFloatTensor *tensor_,
   THLongTensor *offset_, THLongTensor *spatialSize_, float threshold);
 void scn_5_batchAddSample(void **m);
 void scn_5_createMetadataForDenseToSparse(
-  void **m, THLongTensor *spatialSize_, THLongTensor *pad, THLongTensor *nz,
-  long batchSize);
+  void **m, THLongTensor *spatialSize_, THLongTensor *nz, long batchSize);
 void scn_5_freeMetadata(void **metadata);
 void scn_5_generateRuleBooks3s2(void **m);
 void scn_5_generateRuleBooks2s2(void **m);
@@ -93,13 +92,13 @@ void scn_5_setInputSpatialLocations(void **m, THFloatTensor *features,
   THLongTensor *locations, THFloatTensor *vecs, _Bool overwrite);
 void scn_5_getSpatialLocations(void **m, THLongTensor *spatialSize,
     THLongTensor *locations);
-  double scn_6_addSampleFromThresholdedTensor(
+
+double scn_6_addSampleFromThresholdedTensor(
   void **m, THFloatTensor *features_, THFloatTensor *tensor_,
   THLongTensor *offset_, THLongTensor *spatialSize_, float threshold);
 void scn_6_batchAddSample(void **m);
 void scn_6_createMetadataForDenseToSparse(
-  void **m, THLongTensor *spatialSize_, THLongTensor *pad, THLongTensor *nz,
-  long batchSize);
+  void **m, THLongTensor *spatialSize_, THLongTensor *nz, long batchSize);
 void scn_6_freeMetadata(void **metadata);
 void scn_6_generateRuleBooks3s2(void **m);
 void scn_6_generateRuleBooks2s2(void **m);
@@ -110,13 +109,13 @@ void scn_6_setInputSpatialLocations(void **m, THFloatTensor *features,
   THLongTensor *locations, THFloatTensor *vecs, _Bool overwrite);
 void scn_6_getSpatialLocations(void **m, THLongTensor *spatialSize,
     THLongTensor *locations);
-  double scn_7_addSampleFromThresholdedTensor(
+
+double scn_7_addSampleFromThresholdedTensor(
   void **m, THFloatTensor *features_, THFloatTensor *tensor_,
   THLongTensor *offset_, THLongTensor *spatialSize_, float threshold);
 void scn_7_batchAddSample(void **m);
 void scn_7_createMetadataForDenseToSparse(
-  void **m, THLongTensor *spatialSize_, THLongTensor *pad, THLongTensor *nz,
-  long batchSize);
+  void **m, THLongTensor *spatialSize_, THLongTensor *nz, long batchSize);
 void scn_7_freeMetadata(void **metadata);
 void scn_7_generateRuleBooks3s2(void **m);
 void scn_7_generateRuleBooks2s2(void **m);
@@ -127,13 +126,13 @@ void scn_7_setInputSpatialLocations(void **m, THFloatTensor *features,
   THLongTensor *locations, THFloatTensor *vecs, _Bool overwrite);
 void scn_7_getSpatialLocations(void **m, THLongTensor *spatialSize,
     THLongTensor *locations);
-  double scn_8_addSampleFromThresholdedTensor(
+
+double scn_8_addSampleFromThresholdedTensor(
   void **m, THFloatTensor *features_, THFloatTensor *tensor_,
   THLongTensor *offset_, THLongTensor *spatialSize_, float threshold);
 void scn_8_batchAddSample(void **m);
 void scn_8_createMetadataForDenseToSparse(
-  void **m, THLongTensor *spatialSize_, THLongTensor *pad, THLongTensor *nz,
-  long batchSize);
+  void **m, THLongTensor *spatialSize_, THLongTensor *nz, long batchSize);
 void scn_8_freeMetadata(void **metadata);
 void scn_8_generateRuleBooks3s2(void **m);
 void scn_8_generateRuleBooks2s2(void **m);
@@ -144,13 +143,13 @@ void scn_8_setInputSpatialLocations(void **m, THFloatTensor *features,
   THLongTensor *locations, THFloatTensor *vecs, _Bool overwrite);
 void scn_8_getSpatialLocations(void **m, THLongTensor *spatialSize,
     THLongTensor *locations);
-  double scn_9_addSampleFromThresholdedTensor(
+
+double scn_9_addSampleFromThresholdedTensor(
   void **m, THFloatTensor *features_, THFloatTensor *tensor_,
   THLongTensor *offset_, THLongTensor *spatialSize_, float threshold);
 void scn_9_batchAddSample(void **m);
 void scn_9_createMetadataForDenseToSparse(
-  void **m, THLongTensor *spatialSize_, THLongTensor *pad, THLongTensor *nz,
-  long batchSize);
+  void **m, THLongTensor *spatialSize_, THLongTensor *nz, long batchSize);
 void scn_9_freeMetadata(void **metadata);
 void scn_9_generateRuleBooks3s2(void **m);
 void scn_9_generateRuleBooks2s2(void **m);
@@ -161,13 +160,13 @@ void scn_9_setInputSpatialLocations(void **m, THFloatTensor *features,
   THLongTensor *locations, THFloatTensor *vecs, _Bool overwrite);
 void scn_9_getSpatialLocations(void **m, THLongTensor *spatialSize,
     THLongTensor *locations);
-  double scn_10_addSampleFromThresholdedTensor(
+
+double scn_10_addSampleFromThresholdedTensor(
   void **m, THFloatTensor *features_, THFloatTensor *tensor_,
   THLongTensor *offset_, THLongTensor *spatialSize_, float threshold);
 void scn_10_batchAddSample(void **m);
 void scn_10_createMetadataForDenseToSparse(
-  void **m, THLongTensor *spatialSize_, THLongTensor *pad, THLongTensor *nz,
-  long batchSize);
+  void **m, THLongTensor *spatialSize_, THLongTensor *nz, long batchSize);
 void scn_10_freeMetadata(void **metadata);
 void scn_10_generateRuleBooks3s2(void **m);
 void scn_10_generateRuleBooks2s2(void **m);
@@ -178,7 +177,8 @@ void scn_10_setInputSpatialLocations(void **m, THFloatTensor *features,
   THLongTensor *locations, THFloatTensor *vecs, _Bool overwrite);
 void scn_10_getSpatialLocations(void **m, THLongTensor *spatialSize,
     THLongTensor *locations);
-  void scn_cpu_float_AffineReluTrivialConvolution_updateOutput(
+
+void scn_cpu_float_AffineReluTrivialConvolution_updateOutput(
   THFloatTensor *input_features, THFloatTensor *output_features,
   THFloatTensor *affineWeight, THFloatTensor *affineBias, THFloatTensor *convWeight);
 void scn_cpu_float_AffineReluTrivialConvolution_backward(
@@ -348,7 +348,7 @@ void scn_cpu_float1MaxPooling_updateGradInput(
 // SparseToDense
 void scn_cpu_float1SparseToDense_updateOutput(
   THLongTensor *inputSize, void **m, THFloatTensor *input_features,
-  THFloatTensor *output_features, void *rulesBuffer);
+  THFloatTensor *output_features, void *rulesBuffer, long nPlanes);
 void scn_cpu_float1SparseToDense_updateGradInput(
   THLongTensor *inputSize, void **m, THFloatTensor *input_features,
   THFloatTensor *d_input_features, THFloatTensor *d_output_features,
@@ -425,7 +425,7 @@ void scn_cpu_float2MaxPooling_updateGradInput(
 // SparseToDense
 void scn_cpu_float2SparseToDense_updateOutput(
   THLongTensor *inputSize, void **m, THFloatTensor *input_features,
-  THFloatTensor *output_features, void *rulesBuffer);
+  THFloatTensor *output_features, void *rulesBuffer, long nPlanes);
 void scn_cpu_float2SparseToDense_updateGradInput(
   THLongTensor *inputSize, void **m, THFloatTensor *input_features,
   THFloatTensor *d_input_features, THFloatTensor *d_output_features,
@@ -502,7 +502,7 @@ void scn_cpu_float3MaxPooling_updateGradInput(
 // SparseToDense
 void scn_cpu_float3SparseToDense_updateOutput(
   THLongTensor *inputSize, void **m, THFloatTensor *input_features,
-  THFloatTensor *output_features, void *rulesBuffer);
+  THFloatTensor *output_features, void *rulesBuffer, long nPlanes);
 void scn_cpu_float3SparseToDense_updateGradInput(
   THLongTensor *inputSize, void **m, THFloatTensor *input_features,
   THFloatTensor *d_input_features, THFloatTensor *d_output_features,
@@ -579,7 +579,7 @@ void scn_cpu_float4MaxPooling_updateGradInput(
 // SparseToDense
 void scn_cpu_float4SparseToDense_updateOutput(
   THLongTensor *inputSize, void **m, THFloatTensor *input_features,
-  THFloatTensor *output_features, void *rulesBuffer);
+  THFloatTensor *output_features, void *rulesBuffer, long nPlanes);
 void scn_cpu_float4SparseToDense_updateGradInput(
   THLongTensor *inputSize, void **m, THFloatTensor *input_features,
   THFloatTensor *d_input_features, THFloatTensor *d_output_features,
@@ -656,7 +656,7 @@ void scn_cpu_float5MaxPooling_updateGradInput(
 // SparseToDense
 void scn_cpu_float5SparseToDense_updateOutput(
   THLongTensor *inputSize, void **m, THFloatTensor *input_features,
-  THFloatTensor *output_features, void *rulesBuffer);
+  THFloatTensor *output_features, void *rulesBuffer, long nPlanes);
 void scn_cpu_float5SparseToDense_updateGradInput(
   THLongTensor *inputSize, void **m, THFloatTensor *input_features,
   THFloatTensor *d_input_features, THFloatTensor *d_output_features,
@@ -733,7 +733,7 @@ void scn_cpu_float6MaxPooling_updateGradInput(
 // SparseToDense
 void scn_cpu_float6SparseToDense_updateOutput(
   THLongTensor *inputSize, void **m, THFloatTensor *input_features,
-  THFloatTensor *output_features, void *rulesBuffer);
+  THFloatTensor *output_features, void *rulesBuffer, long nPlanes);
 void scn_cpu_float6SparseToDense_updateGradInput(
   THLongTensor *inputSize, void **m, THFloatTensor *input_features,
   THFloatTensor *d_input_features, THFloatTensor *d_output_features,
@@ -810,7 +810,7 @@ void scn_cpu_float7MaxPooling_updateGradInput(
 // SparseToDense
 void scn_cpu_float7SparseToDense_updateOutput(
   THLongTensor *inputSize, void **m, THFloatTensor *input_features,
-  THFloatTensor *output_features, void *rulesBuffer);
+  THFloatTensor *output_features, void *rulesBuffer, long nPlanes);
 void scn_cpu_float7SparseToDense_updateGradInput(
   THLongTensor *inputSize, void **m, THFloatTensor *input_features,
   THFloatTensor *d_input_features, THFloatTensor *d_output_features,
@@ -887,7 +887,7 @@ void scn_cpu_float8MaxPooling_updateGradInput(
 // SparseToDense
 void scn_cpu_float8SparseToDense_updateOutput(
   THLongTensor *inputSize, void **m, THFloatTensor *input_features,
-  THFloatTensor *output_features, void *rulesBuffer);
+  THFloatTensor *output_features, void *rulesBuffer, long nPlanes);
 void scn_cpu_float8SparseToDense_updateGradInput(
   THLongTensor *inputSize, void **m, THFloatTensor *input_features,
   THFloatTensor *d_input_features, THFloatTensor *d_output_features,
@@ -964,7 +964,7 @@ void scn_cpu_float9MaxPooling_updateGradInput(
 // SparseToDense
 void scn_cpu_float9SparseToDense_updateOutput(
   THLongTensor *inputSize, void **m, THFloatTensor *input_features,
-  THFloatTensor *output_features, void *rulesBuffer);
+  THFloatTensor *output_features, void *rulesBuffer, long nPlanes);
 void scn_cpu_float9SparseToDense_updateGradInput(
   THLongTensor *inputSize, void **m, THFloatTensor *input_features,
   THFloatTensor *d_input_features, THFloatTensor *d_output_features,
@@ -1041,7 +1041,8 @@ void scn_cpu_float10MaxPooling_updateGradInput(
 // SparseToDense
 void scn_cpu_float10SparseToDense_updateOutput(
   THLongTensor *inputSize, void **m, THFloatTensor *input_features,
-  THFloatTensor *output_features, void *rulesBuffer);
+  THFloatTensor *output_features, void *rulesBuffer, long nPlanes);
+
 void scn_cpu_float10SparseToDense_updateGradInput(
   THLongTensor *inputSize, void **m, THFloatTensor *input_features,
   THFloatTensor *d_input_features, THFloatTensor *d_output_features,
@@ -1118,7 +1119,7 @@ void scn_cpu_double1MaxPooling_updateGradInput(
 // SparseToDense
 void scn_cpu_double1SparseToDense_updateOutput(
   THLongTensor *inputSize, void **m, THDoubleTensor *input_features,
-  THDoubleTensor *output_features, void *rulesBuffer);
+  THDoubleTensor *output_features, void *rulesBuffer, long nPlanes);
 void scn_cpu_double1SparseToDense_updateGradInput(
   THLongTensor *inputSize, void **m, THDoubleTensor *input_features,
   THDoubleTensor *d_input_features, THDoubleTensor *d_output_features,
@@ -1195,7 +1196,7 @@ void scn_cpu_double2MaxPooling_updateGradInput(
 // SparseToDense
 void scn_cpu_double2SparseToDense_updateOutput(
   THLongTensor *inputSize, void **m, THDoubleTensor *input_features,
-  THDoubleTensor *output_features, void *rulesBuffer);
+  THDoubleTensor *output_features, void *rulesBuffer, long nPlanes);
 void scn_cpu_double2SparseToDense_updateGradInput(
   THLongTensor *inputSize, void **m, THDoubleTensor *input_features,
   THDoubleTensor *d_input_features, THDoubleTensor *d_output_features,
@@ -1272,7 +1273,7 @@ void scn_cpu_double3MaxPooling_updateGradInput(
 // SparseToDense
 void scn_cpu_double3SparseToDense_updateOutput(
   THLongTensor *inputSize, void **m, THDoubleTensor *input_features,
-  THDoubleTensor *output_features, void *rulesBuffer);
+  THDoubleTensor *output_features, void *rulesBuffer, long nPlanes);
 void scn_cpu_double3SparseToDense_updateGradInput(
   THLongTensor *inputSize, void **m, THDoubleTensor *input_features,
   THDoubleTensor *d_input_features, THDoubleTensor *d_output_features,
@@ -1349,7 +1350,7 @@ void scn_cpu_double4MaxPooling_updateGradInput(
 // SparseToDense
 void scn_cpu_double4SparseToDense_updateOutput(
   THLongTensor *inputSize, void **m, THDoubleTensor *input_features,
-  THDoubleTensor *output_features, void *rulesBuffer);
+  THDoubleTensor *output_features, void *rulesBuffer, long nPlanes);
 void scn_cpu_double4SparseToDense_updateGradInput(
   THLongTensor *inputSize, void **m, THDoubleTensor *input_features,
   THDoubleTensor *d_input_features, THDoubleTensor *d_output_features,
@@ -1426,7 +1427,7 @@ void scn_cpu_double5MaxPooling_updateGradInput(
 // SparseToDense
 void scn_cpu_double5SparseToDense_updateOutput(
   THLongTensor *inputSize, void **m, THDoubleTensor *input_features,
-  THDoubleTensor *output_features, void *rulesBuffer);
+  THDoubleTensor *output_features, void *rulesBuffer, long nPlanes);
 void scn_cpu_double5SparseToDense_updateGradInput(
   THLongTensor *inputSize, void **m, THDoubleTensor *input_features,
   THDoubleTensor *d_input_features, THDoubleTensor *d_output_features,
@@ -1503,7 +1504,7 @@ void scn_cpu_double6MaxPooling_updateGradInput(
 // SparseToDense
 void scn_cpu_double6SparseToDense_updateOutput(
   THLongTensor *inputSize, void **m, THDoubleTensor *input_features,
-  THDoubleTensor *output_features, void *rulesBuffer);
+  THDoubleTensor *output_features, void *rulesBuffer, long nPlanes);
 void scn_cpu_double6SparseToDense_updateGradInput(
   THLongTensor *inputSize, void **m, THDoubleTensor *input_features,
   THDoubleTensor *d_input_features, THDoubleTensor *d_output_features,
@@ -1580,7 +1581,7 @@ void scn_cpu_double7MaxPooling_updateGradInput(
 // SparseToDense
 void scn_cpu_double7SparseToDense_updateOutput(
   THLongTensor *inputSize, void **m, THDoubleTensor *input_features,
-  THDoubleTensor *output_features, void *rulesBuffer);
+  THDoubleTensor *output_features, void *rulesBuffer, long nPlanes);
 void scn_cpu_double7SparseToDense_updateGradInput(
   THLongTensor *inputSize, void **m, THDoubleTensor *input_features,
   THDoubleTensor *d_input_features, THDoubleTensor *d_output_features,
@@ -1657,7 +1658,7 @@ void scn_cpu_double8MaxPooling_updateGradInput(
 // SparseToDense
 void scn_cpu_double8SparseToDense_updateOutput(
   THLongTensor *inputSize, void **m, THDoubleTensor *input_features,
-  THDoubleTensor *output_features, void *rulesBuffer);
+  THDoubleTensor *output_features, void *rulesBuffer, long nPlanes);
 void scn_cpu_double8SparseToDense_updateGradInput(
   THLongTensor *inputSize, void **m, THDoubleTensor *input_features,
   THDoubleTensor *d_input_features, THDoubleTensor *d_output_features,
@@ -1734,7 +1735,7 @@ void scn_cpu_double9MaxPooling_updateGradInput(
 // SparseToDense
 void scn_cpu_double9SparseToDense_updateOutput(
   THLongTensor *inputSize, void **m, THDoubleTensor *input_features,
-  THDoubleTensor *output_features, void *rulesBuffer);
+  THDoubleTensor *output_features, void *rulesBuffer, long nPlanes);
 void scn_cpu_double9SparseToDense_updateGradInput(
   THLongTensor *inputSize, void **m, THDoubleTensor *input_features,
   THDoubleTensor *d_input_features, THDoubleTensor *d_output_features,
@@ -1811,7 +1812,7 @@ void scn_cpu_double10MaxPooling_updateGradInput(
 // SparseToDense
 void scn_cpu_double10SparseToDense_updateOutput(
   THLongTensor *inputSize, void **m, THDoubleTensor *input_features,
-  THDoubleTensor *output_features, void *rulesBuffer);
+  THDoubleTensor *output_features, void *rulesBuffer, long nPlanes);
 void scn_cpu_double10SparseToDense_updateGradInput(
   THLongTensor *inputSize, void **m, THDoubleTensor *input_features,
   THDoubleTensor *d_input_features, THDoubleTensor *d_output_features,

PyTorch/sparseconvnet/SCN/header_gpu.h

@@ -120,7 +120,8 @@ void scn_gpu_float1MaxPooling_updateGradInput(
 // SparseToDense
 void scn_gpu_float1SparseToDense_updateOutput(
   THLongTensor *inputSize, void **m, THCudaTensor *input_features,
-  THCudaTensor *output_features, THCudaIntTensor *rulesBuffer);
+  THCudaTensor *output_features, THCudaIntTensor *rulesBuffer,
+  long nPlanes);
 void scn_gpu_float1SparseToDense_updateGradInput(
   THLongTensor *inputSize, void **m, THCudaTensor *input_features,
   THCudaTensor *d_input_features, THCudaTensor *d_output_features,
@@ -197,7 +198,7 @@ void scn_gpu_float2MaxPooling_updateGradInput(
 // SparseToDense
 void scn_gpu_float2SparseToDense_updateOutput(
   THLongTensor *inputSize, void **m, THCudaTensor *input_features,
-  THCudaTensor *output_features, THCudaIntTensor *rulesBuffer);
+  THCudaTensor *output_features, THCudaIntTensor *rulesBuffer, long nPlanes);
 void scn_gpu_float2SparseToDense_updateGradInput(
   THLongTensor *inputSize, void **m, THCudaTensor *input_features,
   THCudaTensor *d_input_features, THCudaTensor *d_output_features,
@@ -274,7 +275,7 @@ void scn_gpu_float3MaxPooling_updateGradInput(
 // SparseToDense
 void scn_gpu_float3SparseToDense_updateOutput(
   THLongTensor *inputSize, void **m, THCudaTensor *input_features,
-  THCudaTensor *output_features, THCudaIntTensor *rulesBuffer);
+  THCudaTensor *output_features, THCudaIntTensor *rulesBuffer, long nPlanes);
 void scn_gpu_float3SparseToDense_updateGradInput(
   THLongTensor *inputSize, void **m, THCudaTensor *input_features,
   THCudaTensor *d_input_features, THCudaTensor *d_output_features,
@@ -351,7 +352,7 @@ void scn_gpu_float4MaxPooling_updateGradInput(
 // SparseToDense
 void scn_gpu_float4SparseToDense_updateOutput(
   THLongTensor *inputSize, void **m, THCudaTensor *input_features,
-  THCudaTensor *output_features, THCudaIntTensor *rulesBuffer);
+  THCudaTensor *output_features, THCudaIntTensor *rulesBuffer, long nPlanes);
 void scn_gpu_float4SparseToDense_updateGradInput(
   THLongTensor *inputSize, void **m, THCudaTensor *input_features,
   THCudaTensor *d_input_features, THCudaTensor *d_output_features,
@@ -428,7 +429,7 @@ void scn_gpu_float5MaxPooling_updateGradInput(
 // SparseToDense
 void scn_gpu_float5SparseToDense_updateOutput(
   THLongTensor *inputSize, void **m, THCudaTensor *input_features,
-  THCudaTensor *output_features, THCudaIntTensor *rulesBuffer);
+  THCudaTensor *output_features, THCudaIntTensor *rulesBuffer, long nPlanes);
 void scn_gpu_float5SparseToDense_updateGradInput(
   THLongTensor *inputSize, void **m, THCudaTensor *input_features,
   THCudaTensor *d_input_features, THCudaTensor *d_output_features,
@@ -505,7 +506,7 @@ void scn_gpu_float6MaxPooling_updateGradInput(
 // SparseToDense
 void scn_gpu_float6SparseToDense_updateOutput(
   THLongTensor *inputSize, void **m, THCudaTensor *input_features,
-  THCudaTensor *output_features, THCudaIntTensor *rulesBuffer);
+  THCudaTensor *output_features, THCudaIntTensor *rulesBuffer, long nPlanes);
 void scn_gpu_float6SparseToDense_updateGradInput(
   THLongTensor *inputSize, void **m, THCudaTensor *input_features,
   THCudaTensor *d_input_features, THCudaTensor *d_output_features,
@@ -582,7 +583,7 @@ void scn_gpu_float7MaxPooling_updateGradInput(
 // SparseToDense
 void scn_gpu_float7SparseToDense_updateOutput(
   THLongTensor *inputSize, void **m, THCudaTensor *input_features,
-  THCudaTensor *output_features, THCudaIntTensor *rulesBuffer);
+  THCudaTensor *output_features, THCudaIntTensor *rulesBuffer, long nPlanes);
 void scn_gpu_float7SparseToDense_updateGradInput(
   THLongTensor *inputSize, void **m, THCudaTensor *input_features,
   THCudaTensor *d_input_features, THCudaTensor *d_output_features,
@@ -659,7 +660,7 @@ void scn_gpu_float8MaxPooling_updateGradInput(
 // SparseToDense
 void scn_gpu_float8SparseToDense_updateOutput(
   THLongTensor *inputSize, void **m, THCudaTensor *input_features,
-  THCudaTensor *output_features, THCudaIntTensor *rulesBuffer);
+  THCudaTensor *output_features, THCudaIntTensor *rulesBuffer, long nPlanes);
 void scn_gpu_float8SparseToDense_updateGradInput(
   THLongTensor *inputSize, void **m, THCudaTensor *input_features,
   THCudaTensor *d_input_features, THCudaTensor *d_output_features,
@@ -736,7 +737,7 @@ void scn_gpu_float9MaxPooling_updateGradInput(
 // SparseToDense
 void scn_gpu_float9SparseToDense_updateOutput(
   THLongTensor *inputSize, void **m, THCudaTensor *input_features,
-  THCudaTensor *output_features, THCudaIntTensor *rulesBuffer);
+  THCudaTensor *output_features, THCudaIntTensor *rulesBuffer, long nPlanes);
 void scn_gpu_float9SparseToDense_updateGradInput(
   THLongTensor *inputSize, void **m, THCudaTensor *input_features,
   THCudaTensor *d_input_features, THCudaTensor *d_output_features,
@@ -813,7 +814,7 @@ void scn_gpu_float10MaxPooling_updateGradInput(
 // SparseToDense
 void scn_gpu_float10SparseToDense_updateOutput(
   THLongTensor *inputSize, void **m, THCudaTensor *input_features,
-  THCudaTensor *output_features, THCudaIntTensor *rulesBuffer);
+  THCudaTensor *output_features, THCudaIntTensor *rulesBuffer, long nPlanes);
 void scn_gpu_float10SparseToDense_updateGradInput(
   THLongTensor *inputSize, void **m, THCudaTensor *input_features,
   THCudaTensor *d_input_features, THCudaTensor *d_output_features,