Split operation improvement (#262)

* libnd4j moved split operation implementation to helpers before special case adding

Signed-off-by: Oleg <oleg.semeniv@gmail.com>

* libnd4j minor fixes for general split operation move, merge master

Signed-off-by: Oleg <oleg.semeniv@gmail.com>

* libndj4 split cpu implementation

Signed-off-by: Oleg <oleg.semeniv@gmail.com>

* - provide cuda helper for split op

Signed-off-by: Yurii <iuriish@yahoo.com>

* - minor correction

Signed-off-by: Yurii <iuriish@yahoo.com>

* - minor correction 2

Signed-off-by: Yurii <iuriish@yahoo.com>

Co-authored-by: Yurii Shyrma <iuriish@yahoo.com>
Co-authored-by: raver119 <raver119@gmail.com>

libnd4j/include/ops/declarable/generic/parity_ops/split.cpp

@@ -22,6 +22,7 @@
 #if NOT_EXCLUDED(OP_split)
 
 #include <ops/declarable/headers/parity_ops.h>
+#include<ops/declarable/helpers/transforms.h>
 #include <array>
 
 namespace nd4j {
@@ -65,30 +66,13 @@ namespace ops {
 
         REQUIRE_TRUE(input->sizeAt(axis) % num_splits == 0, 0, "Split: num_splits has wrong value, remainder of division should be 0, but it's %i", input->sizeAt(axis) % num_splits);
 
-        int pos = 0;
+        std::vector<NDArray*> outArrs(num_splits);
-        int split = input->sizeAt(axis) / num_splits;
-        std::vector<Nd4jLong> indices(2 * input->rankOf());
-        
         for (int e = 0; e < num_splits; e++) {
-            
+            outArrs[e] = OUTPUT_VARIABLE(e);
-            auto out = OUTPUT_VARIABLE(e);
-            
-            for (int d = 0; d < input->rankOf(); d++) {
-                if (d == axis) {
-                    indices[2*d]     = pos;
-                    indices[2*d + 1] = pos + split; 
-                }
-                else 
-                    indices[2*d] = indices[2*d + 1] = 0;
-            }
-
-            auto sub = (*input)(indices, true);
-            
-            out->assign(sub);
-
-            pos += split;
         }
 
+        helpers::split(block.launchContext(), *input, outArrs, axis);
+
         return Status::OK();
     }
 

libnd4j/include/ops/declarable/helpers/cpu/split.cpp

@@ -0,0 +1,43 @@
+/*******************************************************************************
+ * Copyright (c) 2019-2020 Konduit K.K.
+ *
+ * This program and the accompanying materials are made available under the
+ * terms of the Apache License, Version 2.0 which is available at
+ * https://www.apache.org/licenses/LICENSE-2.0.
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
+ * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
+ * License for the specific language governing permissions and limitations
+ * under the License.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ ******************************************************************************/
+
+ //
+ //  @author Oleh Semeniv (oleg.semeniv@gmail.com)
+ //
+
+
+#include <ops/declarable/helpers/transforms.h>
+#include <ops/specials.h>
+
+namespace nd4j {
+namespace ops {
+namespace helpers {
+
+//////////////////////////////////////////////////////////////////////////
+template<typename T>
+static void split_(const NDArray& input, const std::vector<NDArray*>& outArrs, const int axis) {
+    nd4j::SpecialMethods<T>::splitCpuGeneric(input, outArrs, axis);
+}
+
+void split(nd4j::LaunchContext* context, const NDArray& input, std::vector<NDArray*>& outArrs, const int axis) {
+    BUILD_SINGLE_SELECTOR(input.dataType(), split_, (input, outArrs, axis), LIBND4J_TYPES);
+}
+
+BUILD_SINGLE_TEMPLATE(template void split_, (const NDArray& input, const std::vector<NDArray*>& outArrs, const int axis), LIBND4J_TYPES);
+
+}
+}
+}

libnd4j/include/ops/declarable/helpers/cuda/split.cu

@@ -0,0 +1,187 @@
+/*******************************************************************************
+ * Copyright (c) 2015-2018 Skymind, Inc.
+ * Copyright (c) 2019 Konduit K.K.
+ *
+ * This program and the accompanying materials are made available under the
+ * terms of the Apache License, Version 2.0 which is available at
+ * https://www.apache.org/licenses/LICENSE-2.0.
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
+ * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
+ * License for the specific language governing permissions and limitations
+ * under the License.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ ******************************************************************************/
+
+//
+// @author Yurii Shyrma (iuriish@yahoo.com)
+//
+
+
+#include<ops/declarable/helpers/transforms.h>
+#include <array/ResultSet.h>
+#include <helpers/ShapeUtils.h>
+#include <numeric>
+#include <NDArrayFactory.h>
+#include <helpers/TAD.h>
+#include <exceptions/cuda_exception.h>
+#include <PointersManager.h>
+#include <ConstantTadHelper.h>
+
+namespace nd4j    {
+namespace ops     {
+namespace helpers {
+
+
+///////////////////////////////////////////////////////////////////
+template<typename T>
+__global__ static void splitCuda(const void* vx, const Nd4jLong* xShapeInfo, void* pVz, const Nd4jLong* zTadShapeInfo, const int axis) {
+
+    const T* x = reinterpret_cast<const T*>(vx);
+
+    __shared__ Nd4jLong xLen, totalThreads;
+    __shared__ int xRank, zDim;
+
+    if (threadIdx.x == 0) {
+        xLen  = shape::length(xShapeInfo);
+        xRank = shape::rank(xShapeInfo);
+        zDim  = shape::shapeOf(zTadShapeInfo)[axis];          // same for all input arrays
+        totalThreads = gridDim.z * blockDim.z;
+    }
+    __syncthreads();
+
+    const auto tid = blockIdx.z * blockDim.z + threadIdx.z;
+
+    Nd4jLong coords[MAX_RANK];
+
+    for (uint64_t i = tid; i < xLen; i += totalThreads) {
+
+        shape::index2coords(i, xShapeInfo, coords);
+
+        const auto xOffset = shape::getOffset(xShapeInfo, coords);
+
+        auto *z = reinterpret_cast<T*>(reinterpret_cast<void **>(pVz)[coords[axis] / zDim]);
+
+        coords[axis] %= zDim;
+
+        const auto zOffset = shape::getOffset(zTadShapeInfo, coords);
+
+        z[zOffset] = x[xOffset];
+    }
+}
+
+///////////////////////////////////////////////////////////////////
+template<typename T>
+__host__ static void splitCudaLauncher(const int blocksPerGrid, const int threadsPerBlock, const cudaStream_t *stream,
+                                       const void* vx, const Nd4jLong* xShapeInfo, void* pVz, const Nd4jLong* zTadShapeInfo, const int axis) {
+
+    splitCuda<T><<<blocksPerGrid, threadsPerBlock, 256, *stream>>>(vx, xShapeInfo, pVz, zTadShapeInfo, axis);
+}
+BUILD_SINGLE_TEMPLATE(template void splitCudaLauncher, (const int blocksPerGrid, const int threadsPerBlock, const cudaStream_t *stream, const void* vx, const Nd4jLong* xShapeInfo, void* pVz, const Nd4jLong* zTadShapeInfo, const int axis), LIBND4J_TYPES);
+
+//////////////////////////////////////////////////////////////////////////
+void split(nd4j::LaunchContext* context, const NDArray& input, std::vector<NDArray*>& outArrs, const int axis) {
+
+    const int numOfSubArrs = outArrs.size();
+    const auto sizeofT     = input.sizeOfT();
+
+    for(int i = 0; i < numOfSubArrs; ++i)
+        outArrs[i]->syncToDevice();
+    input.syncToDevice();
+
+    bool luckCase1 = ((axis == 0 && input.ordering() == 'c') || (axis == input.rankOf() - 1 && input.ordering() == 'f')) && input.ews() == 1;
+
+    if(luckCase1) {
+        for (uint i = 0; i < numOfSubArrs; ++i) {
+            luckCase1 &= outArrs[i]->ordering() == input.ordering() && outArrs[i]->ews() == 1;
+            if(!luckCase1)
+                break;
+        }
+    }
+
+    if(luckCase1) {     // for example {1,10} + {2,10} + {3,10} = {6, 10} order c; or {10,1} + {10,2} + {10,3} = {10, 6} order f
+
+        void* x = static_cast<int8_t*>(input.getSpecialBuffer());
+
+        for (uint i = 0; i < numOfSubArrs; ++i) {
+            const auto memAmountToCopy = outArrs[i]->lengthOf() * sizeofT;
+            cudaMemcpyAsync(static_cast<int8_t*>(outArrs[i]->getSpecialBuffer()), x, memAmountToCopy, cudaMemcpyDeviceToDevice, *context->getCudaStream());
+            x = static_cast<int8_t*>(x) + memAmountToCopy;
+        }
+
+        if(cudaStreamSynchronize(*context->getCudaStream()) != 0)
+            throw std::runtime_error("split cuda: luckCase1 failed!");
+
+        for(int i = 0; i < numOfSubArrs; ++i)
+            outArrs[i]->tickWriteDevice();
+        input.tickReadDevice();
+
+        return;
+    }
+
+    const bool isXcontin = input.strideAt(axis) == 1;
+    bool areOutputsContin = true;
+    bool allSameOrder    = true;
+
+    if(isXcontin) {
+        for (uint i = 0; i < outArrs.size(); ++i) {
+            areOutputsContin &= outArrs[i]->strideAt(axis) == 1;
+            allSameOrder     &= input.ordering() == outArrs[i]->ordering();
+            if(!areOutputsContin || !allSameOrder)
+                break;
+        }
+    }
+
+    const bool luckCase2 = isXcontin && areOutputsContin && allSameOrder;
+
+    if(luckCase2) {     // for example {2,1,3} + {2,5,3} + {2,10,3} = {2,16,3}, here axis 1 shoud have stride = 1 for all inputs arrays and input array
+
+        const auto xDim = input.sizeAt(axis);
+        const auto zDim = outArrs[0]->sizeAt(axis);     // same for all outArrs
+
+        for (uint i = 0; i < input.lengthOf() / xDim; ++i) {
+
+            const auto iShift = i * sizeofT;
+            void* x = static_cast<int8_t*>(input.getSpecialBuffer()) + xDim * iShift;
+
+            for (uint j = 0; j < numOfSubArrs; ++j) {
+                void* z = static_cast<int8_t*>(outArrs[j]->getSpecialBuffer()) + zDim * iShift;
+                const auto memSizeToCopy = zDim * sizeofT;
+                cudaMemcpyAsync(z, x, memSizeToCopy, cudaMemcpyDeviceToDevice, *context->getCudaStream());
+                x = static_cast<int8_t*>(x) + memSizeToCopy;
+            }
+        }
+
+        if(cudaStreamSynchronize(*context->getCudaStream()) != 0)
+            throw std::runtime_error("split cuda: luckCase2 failed!");
+    }
+    else {      // general (slower) case
+
+        const int threadsPerBlock = MAX_NUM_THREADS / 2;
+        const int blocksPerGrid = (input.lengthOf() + threadsPerBlock - 1) / threadsPerBlock;
+
+        // prepare arrays of pointers on buffers and shapes
+        std::vector<void*> hOutBuffers(numOfSubArrs);
+
+        for(int i = 0; i < numOfSubArrs; ++i)
+            hOutBuffers[i]   = outArrs[i]->getSpecialBuffer();
+
+        PointersManager manager(context, "helpers::split");
+
+        void* dOutBuffers = manager.replicatePointer(hOutBuffers.data(), hOutBuffers.size() * sizeof(void*));
+
+        BUILD_SINGLE_SELECTOR(input.dataType(), splitCudaLauncher, (blocksPerGrid, threadsPerBlock, context->getCudaStream(), input.getSpecialBuffer(), input.getSpecialShapeInfo(), dOutBuffers, outArrs[0]->specialShapeInfo(), axis), LIBND4J_TYPES);
+
+        manager.synchronize();
+    }
+
+    for(int i = 0; i < numOfSubArrs; ++i)
+        outArrs[i]->tickWriteDevice();
+    input.tickReadDevice();
+}
+
+}
+}
+}

libnd4j/include/ops/declarable/helpers/transforms.h

@@ -73,6 +73,8 @@ namespace helpers {
 	void concat(nd4j::LaunchContext * context, const std::vector<NDArray*>& inArrs, NDArray& output, const int axis);
 
 	void tileBP(nd4j::LaunchContext * context, const NDArray& gradO /*input*/, NDArray& gradI /*output*/, const std::vector<Nd4jLong> reps);
+
+	void split(nd4j::LaunchContext* context, const NDArray& input, std::vector<NDArray*>& outArrs, const int axis);
 }
 }
 }

libnd4j/include/ops/impl/specials_single.hpp

@@ -218,6 +218,99 @@ void SpecialMethods<T>::concatCpuGeneric(int dimension, int numArrays, Nd4jPoint
 }
 
 
+template <typename T>
+void SpecialMethods<T>::splitCpuGeneric(const NDArray& input, const std::vector<NDArray*>& outArrs, const int axis) {
+
+    int numSplits = outArrs.size();
+
+    const auto sizeofT = input.sizeOfT();
+
+    T* xBuff = input.bufferAsT<T>();
+
+    bool luckCase1 = ((axis == 0 && input.ordering() == 'c') || (axis == input.rankOf() - 1 && input.ordering() == 'f')) && input.ews() == 1;
+
+    if (luckCase1) {
+        for (uint i = 0; i < numSplits; ++i) {
+            luckCase1 &= outArrs[i]->ordering() == input.ordering() && outArrs[i]->ews() == 1;
+            if (!luckCase1)
+                break;
+        }
+    }
+
+    if (luckCase1) {
+
+        T* x = const_cast<T*>(xBuff);
+        for (uint i = 0; i < numSplits; ++i) {
+            const auto memAmountToCopy = outArrs[i]->lengthOf();
+            memcpy(outArrs[i]->bufferAsT<T>(), x, memAmountToCopy * sizeofT);
+            x += memAmountToCopy;
+        }
+        return;
+    }
+
+    const bool isXcontin = input.strideAt(axis) == 1 && input.ordering() == 'c';
+    bool areOutsContin = true;
+    bool allSameOrder = true;
+
+    if (isXcontin) {
+        for (uint i = 0; i < numSplits; ++i) {
+            areOutsContin &= outArrs[i]->strideAt(axis) == 1;
+            allSameOrder &= outArrs[i]->ordering() == input.ordering();
+            if (!areOutsContin || !allSameOrder)
+                break;
+        }
+    }
+
+    const bool luckCase2 = isXcontin && areOutsContin && allSameOrder;
+
+    if (luckCase2) {
+
+        const uint xDim = input.sizeAt(axis);
+
+        for (uint i = 0; i < input.lengthOf() / xDim; ++i) {
+
+            T* x = xBuff + xDim * i;
+
+            for (uint j = 0; j < numSplits; ++j) {
+                const auto zDim = outArrs[j]->sizeAt(axis);
+                T* z = outArrs[j]->bufferAsT<T>() + zDim * i;
+                memcpy(z, x, zDim * sizeofT);
+                z += zDim;
+                x += zDim;
+            }
+        }
+
+        return;
+    }
+
+    uint zDim = outArrs[0]->sizeAt(axis);
+    // general case
+
+    auto func = PRAGMA_THREADS_FOR{
+
+        Nd4jLong coords[MAX_RANK];
+        for (auto i = start; i < stop; i += increment) {
+
+            shape::index2coords(i, input.getShapeInfo(), coords);
+            const auto xOffset = shape::getOffset(input.getShapeInfo(), coords);
+
+            uint outArrIdx = 0;
+
+            while (coords[axis] >= zDim) {
+                coords[axis] -= zDim;
+                ++outArrIdx;
+            }
+
+            T* z = outArrs[outArrIdx]->bufferAsT<T>();
+            const auto zOffset = shape::getOffset(outArrs[outArrIdx]->getShapeInfo(), coords);
+            z[zOffset] = xBuff[xOffset];
+        }
+    };
+
+    samediff::Threads::parallel_for(func, 0, input.lengthOf());
+}
+
+
 /**
  * This kernel accumulates X arrays, and stores result into Z
  *

libnd4j/include/ops/specials.h

@@ -67,6 +67,8 @@ namespace nd4j {
 
         static void decodeBitmapGeneric(void *dx, Nd4jLong N, void *dz, Nd4jLong *zShapeInfo);
         static Nd4jLong encodeBitmapGeneric(void *dx, Nd4jLong *zShapeInfo, Nd4jLong N, int *dz, float threshold);
+
+        static void splitCpuGeneric(const NDArray& input, const std::vector<NDArray*>& outArrs, const int axis);
     };
 
     template <typename X, typename Y>