Merge pull request #863 from reyoung/feature/clean_sparse_row_matrix

Extract RowBuffer class for SparseRowMatrix.

paddle/math/RowBuffer.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://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. */
+
+#pragma once
+#include <vector>
+#include "MemoryHandle.h"
+#include "paddle/utils/Util.h"
+
+namespace paddle {
+
+/**
+ * @brief The RowBuffer class
+ * Represent the SparseRow Matrix Data.
+ *
+ * If not set memory handler, then the data could be auto growth.
+ */
+class RowBuffer {
+public:
+  /**
+   * @brief RowBuffer create a auto-growth row buffer. The row length is width.
+   * @param width the length of each row, a.k.a matrix width.
+   */
+  explicit RowBuffer(size_t width) : width_(width) {}
+
+  /**
+   * @brief RowBuffer create a row buffer, which cannot be auto-growth.
+   * @param mem the pre-allocated memory.
+   * @param width the length of each row, a.k.a matrix width.
+   */
+  RowBuffer(const CpuMemHandlePtr& mem, size_t width)
+      : preallocatedBuf_(mem), width_(width) {}
+
+  /**
+   * @brief resize resize the buffer with rowCount
+   * @param rowCnt number of row. matrix height.
+   */
+  inline void resize(int rowCnt) {
+    if (preallocatedBuf_) {
+      CHECK(preallocatedBuf_->getSize() >= rowCnt * width_ * sizeof(real));
+    } else {
+      rowStore_.resize(rowCnt * width_);
+    }
+  }
+
+  /**
+   * @brief get a row buffer with row index.
+   * @param row the index of row.
+   * @return row buffer.
+   */
+  inline real* get(int row) const {
+    if (preallocatedBuf_) {
+      CHECK_LE((row + 1) * width_ * sizeof(real), preallocatedBuf_->getSize());
+      return reinterpret_cast<real*>(preallocatedBuf_->getBuf()) + row * width_;
+    } else {
+      CHECK_LE((row + 1) * width_, rowStore_.size());
+      return const_cast<real*>(rowStore_.data() + row * width_);
+    }
+  }
+
+  /**
+   * @brief get a row buffer with row index. If row index is larger than local
+   *        buffer, the size of local buffer will grow.
+   * @param row the index of row.
+   * @return row buffer.
+   */
+  inline real* getWithAutoGrowth(int row) {
+    if (preallocatedBuf_) {
+      return get(row);
+    } else {
+      if ((rowStore_.size() <= row * width_)) {
+        rowStore_.resize((row + 1) * width_);
+      }
+      return rowStore_.data() + row * width_;
+    }
+  }
+
+  /**
+   * @return raw data buffer.
+   */
+  inline real* data() {
+    if (preallocatedBuf_) {
+      return reinterpret_cast<real*>(preallocatedBuf_->getBuf());
+    } else {
+      return rowStore_.data();
+    }
+  }
+
+  /**
+   * @brief clear local buffer. It only affect auto-growth buffer.
+   */
+  inline void clear() { rowStore_.clear(); }
+
+  /**
+   * @brief get current number of rows.
+   * @return number of rows.
+   */
+  inline size_t getRowCount() const {
+    if (preallocatedBuf_) {
+      return preallocatedBuf_->getSize() / sizeof(real) / width_;
+    } else {
+      return rowStore_.size() / width_;
+    }
+  }
+
+  /**
+   * @brief get is this buffer can automatically grow or not.
+   * @return ture if can automacitally grow.
+   */
+  inline bool isAutoGrowth() const { return !preallocatedBuf_; }
+
+  /**
+   * @brief return the width of matrix. a.k.a length of row.
+   * @return width of matrix
+   */
+  inline size_t getWidth() const { return width_; }
+
+private:
+  //! TODO(yuyang18): Add resize method to CpuMemHandlePtr, then we can get rid
+  //! of std::vector here.
+  CpuMemHandlePtr preallocatedBuf_;
+  std::vector<real, AlignedAllocator<real, 32>> rowStore_;
+  size_t width_;
+};
+}  // namespace paddle

paddle/math/SparseRowMatrix.h

@@ -18,6 +18,7 @@ limitations under the License. */
 #include <string.h>
 #include <algorithm>
 #include "Matrix.h"
+#include "RowBuffer.h"
 #include "paddle/utils/Util.h"
 
 DECLARE_bool(allow_inefficient_sparse_update);
@@ -45,12 +46,9 @@ public:
                      IndexDictPtr indexDictHandle = nullptr,
                      bool trans = false)
       : CpuMatrix(nullptr, height, width, trans),
-        storeMat_(dataHandle,
-                  dataHandle ? dataHandle->getSize() / sizeof(real) / width : 0,
-                  width,
-                  trans),
         indexDictHandle_(indexDictHandle) {
     init(height, width);
+    buf_.reset(new RowBuffer(dataHandle, width));
   }
 
   virtual ~SparseRowCpuMatrix() {}
@@ -71,25 +69,16 @@ public:
    *
    *  @param row row id in local storage
    */
-  real* getLocalRow(size_t row) {
-    if (storeMat_.getData()) return storeMat_.rowBuf(row);
-    if (rowStore_.size() <= row * width_) {
-      rowStore_.resize((row + 1) * width_);
-    }
-    return rowStore_.data() + row * width_;
-  }
+  real* getLocalRow(size_t row) { return buf_->getWithAutoGrowth(row); }
 
   /**
-   *  reserve the storage for rows according to current size of indexDictHandle.
+   *  reserve the storage for rows according to current size of
+   * indexDictHandle.
    *
    *  This is only used when SparseRowCpuMatrix is constructed with
    *  indexDictHandle.
    */
-  void reserveStore() {
-    if (!storeMat_.getData() && !localIndices_->empty()) {
-      rowStore_.resize(localIndices_->size() * width_);
-    }
-  }
+  void reserveStore() { buf_->resize(localIndices_->size()); }
 
   // row is the row id in the original matrix
   virtual real* getRowBuf(size_t row) { return getRow(row); }
@@ -117,7 +106,8 @@ public:
    *
    * If L1 decay set use L1, else if L2 set use L2, otherwise no decay atall.
    *
-   * t0 is a int vector used by L1/L2 decay, size = height of parameter matrix,
+   * t0 is a int vector used by L1/L2 decay, size = height of parameter
+   * matrix,
    * store the time that each weight row last updated.
    *
    * Time is batchId, currentTime is current batchId.
@@ -176,8 +166,7 @@ public:
 protected:
   template <typename Func>
   void apply(Func f) {
-    real* data = storeMat_.getData() ? storeMat_.getData() : rowStore_.data();
-    f(data, localIndices_->size() * width_);
+    f(buf_->data(), localIndices_->size() * width_);
   }
 
   void init(size_t height, size_t width);
@@ -188,25 +177,24 @@ protected:
       globalIndices_[id] = kUnusedId_;
     }
     localIndices_->clear();
-    rowStore_.clear();
+    buf_->clear();
   }
 
   inline void checkStoreSize() {
-    if (storeMat_.getData()) {
-      CHECK_LE(localIndices_->size(), storeMat_.getHeight());
-    } else if (!FLAGS_allow_inefficient_sparse_update) {
-      if (localIndices_->size() > 0.5 * height_) {
+    if (buf_->isAutoGrowth()) {
+      if (buf_->getRowCount() > 0.5 * height_) {
         LOG(WARNING)
             << "There are more than 0.5*height (" << localIndices_->size()
             << ") rows are used for sparse "
             << "update, which is not efficient. Considering not use "
             << "sparse_update or set --allow_inefficient_sparse_update=true";
       }
+    } else {
+      CHECK_LE(localIndices_->size(), buf_->getRowCount());
     }
   }
 
-  CpuMatrix storeMat_;
-  std::vector<real, AlignedAllocator<real, 32>> rowStore_;
+  std::unique_ptr<RowBuffer> buf_;
   IndexDictPtr indexDictHandle_;
   std::vector<unsigned int>* localIndices_;  // =&indexDictHandle_->localIndices
   unsigned int* globalIndices_;  // =indexDictHandle_->globalIndices.data();

paddle/math/tests/CMakeLists.txt

@@ -4,6 +4,7 @@ add_simple_unittest(test_ExecViaCpu)
 add_simple_unittest(test_SIMDFunctions)
 add_simple_unittest(test_TrainingAlgorithm)
 add_simple_unittest(test_SparseMatrix)
+add_simple_unittest(test_RowBuffer)
 
 # TODO(yuyang18): Refactor TestUtil.cpp. Remove this cross module reference.
 add_unittest(test_matrixCompare

paddle/math/tests/test_RowBuffer.cpp

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://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. */
+
+#include <gtest/gtest.h>
+#include "paddle/math/RowBuffer.h"
+
+TEST(RowBuffer, testAutoGrow) {
+  paddle::RowBuffer buf(128);
+  ASSERT_EQ(128, buf.getWidth());
+  ASSERT_TRUE(buf.isAutoGrowth());
+  buf.resize(2);
+  ASSERT_EQ(2, buf.getRowCount());
+  for (size_t i = 0; i < buf.getWidth() * 2; ++i) {
+    buf.data()[i] = i;
+  }
+  for (size_t i = 0; i < buf.getRowCount(); ++i) {
+    for (size_t j = 0; j < buf.getWidth(); ++j) {
+      ASSERT_NEAR(i * buf.getWidth() + j, buf.get(i)[j], 1e-5);
+    }
+  }
+
+  auto data = buf.getWithAutoGrowth(2);
+  for (size_t i = 0; i < buf.getWidth(); ++i) {
+    data[i] = i;
+  }
+
+  ASSERT_EQ(3, buf.getRowCount());
+  for (size_t i = 0; i < buf.getRowCount() - 1; ++i) {
+    for (size_t j = 0; j < buf.getWidth(); ++j) {
+      ASSERT_NEAR(i * buf.getWidth() + j, buf.get(i)[j], 1e-5);
+    }
+  }
+  for (size_t i = 0; i < buf.getWidth(); ++i) {
+    ASSERT_NEAR(i, buf.get(2)[i], 1e-5);
+  }
+}
+
+TEST(RowBuffer, testWithMemBuf) {
+  paddle::CpuMemHandlePtr mem =
+      std::make_shared<paddle::CpuMemoryHandle>(128 * 2 * sizeof(real));
+  paddle::RowBuffer buf(mem, 128);
+  ASSERT_TRUE(!buf.isAutoGrowth());
+  ASSERT_EQ(2, buf.getRowCount());
+  for (size_t i = 0; i < buf.getWidth() * 2; ++i) {
+    buf.data()[i] = i;
+  }
+  for (size_t i = 0; i < buf.getRowCount(); ++i) {
+    for (size_t j = 0; j < buf.getWidth(); ++j) {
+      ASSERT_NEAR(i * buf.getWidth() + j, buf.getWithAutoGrowth(i)[j], 1e-5);
+    }
+  }
+
+  ASSERT_DEATH_IF_SUPPORTED(buf.getWithAutoGrowth(3), ".*");
+}