Merge pull request #4590 from wangkuiyi/paddle_only_cpu

Add -D PADDLE_WITH_CUDA in cmake/configure.cmake

cmake/configure.cmake

@@ -53,7 +53,8 @@ if(NOT WITH_GPU)
 
     list(APPEND CMAKE_CXX_SOURCE_FILE_EXTENSIONS cu)
 else()
-    add_definitions(-DPADDLE_WITH_GPU)
+    add_definitions(-DPADDLE_WITH_CUDA)
+
     FIND_PACKAGE(CUDA REQUIRED)
 
     if(${CUDA_VERSION_MAJOR} VERSION_LESS 7)

paddle/api/Util.cpp

@@ -47,7 +47,7 @@ bool isUsingGpu() { return FLAGS_use_gpu; }
 void setUseGpu(bool useGpu) { FLAGS_use_gpu = useGpu; }
 
 bool isGpuVersion() {
-#ifndef PADDLE_WITH_GPU
+#ifndef PADDLE_WITH_CUDA
   return false;
 #else
   return true;

paddle/capi/Matrix.cpp

@@ -46,7 +46,7 @@ paddle_error paddle_matrix_set_row(paddle_matrix mat,
   if (rowID >= ptr->mat->getHeight()) return kPD_OUT_OF_RANGE;
   paddle::real* buf = ptr->mat->getRowBuf(rowID);
   size_t width = ptr->mat->getWidth();
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
   hl_memcpy(buf, rowArray, sizeof(paddle::real) * width);
 #else
   std::copy(rowArray, rowArray + width, buf);

paddle/framework/lod_tensor.h

@@ -15,7 +15,7 @@
 #pragma once
 
 #include <memory>
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 #include <thrust/device_vector.h>
 #include <thrust/host_vector.h>
 #include <thrust/system/cuda/experimental/pinned_allocator.h>
@@ -29,7 +29,7 @@
 namespace paddle {
 namespace framework {
 
-#ifndef PADDLE_WITH_GPU
+#ifndef PADDLE_WITH_CUDA
 template <typename T>
 using Vector = std::vector<T>;
 #else

paddle/framework/op_registry.h

@@ -211,7 +211,7 @@ class OpKernelRegistrar : public Registrar {
 // TODO(fengjiayi): The following macros
 // seems ugly, do we have better method?
 
-#ifndef PADDLE_WITH_GPU
+#ifndef PADDLE_WITH_CUDA
 #define USE_OP_KERNEL(op_type) USE_OP_DEVICE_KERNEL(op_type, CPU)
 #else
 #define USE_OP_KERNEL(op_type)        \

paddle/framework/operator.cc

@@ -25,7 +25,7 @@ Eigen::DefaultDevice& ExecutionContext::GetEigenDevice<
   return *device_context_.GetEigenDevice<platform::CPUPlace>();
 }
 
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 template <>
 Eigen::GpuDevice&
 ExecutionContext::GetEigenDevice<platform::GPUPlace, Eigen::GpuDevice>() const {

paddle/framework/tensor_impl.h

@@ -65,7 +65,7 @@ inline T* Tensor::mutable_data(platform::Place place) {
       holder_.reset(new PlaceholderImpl<T, platform::CPUPlace>(
           boost::get<platform::CPUPlace>(place), size));
     } else if (platform::is_gpu_place(place)) {
-#ifndef PADDLE_WITH_GPU
+#ifndef PADDLE_WITH_CUDA
       PADDLE_THROW("'GPUPlace' is not supported in CPU only device.");
     }
 #else
@@ -103,7 +103,7 @@ inline void Tensor::CopyFrom(const Tensor& src,
     memory::Copy(boost::get<platform::CPUPlace>(dst_place), dst_ptr,
                  boost::get<platform::CPUPlace>(src_place), src_ptr, size);
   }
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
   else if (platform::is_gpu_place(src_place) &&
            platform::is_cpu_place(dst_place)) {
     memory::Copy(boost::get<platform::CPUPlace>(dst_place), dst_ptr,

paddle/framework/tensor_test.cc

@@ -74,7 +74,7 @@ TEST(Tensor, MutableData) {
     EXPECT_EQ(p1, p2);
   }
 
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
   {
     Tensor src_tensor;
     float* p1 = nullptr;
@@ -126,7 +126,7 @@ TEST(Tensor, ShareDataWith) {
     ASSERT_EQ(src_tensor.data<int>(), dst_tensor.data<int>());
   }
 
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
   {
     Tensor src_tensor;
     Tensor dst_tensor;
@@ -163,7 +163,7 @@ TEST(Tensor, Slice) {
     EXPECT_EQ(src_data_address + 3 * 4 * 1 * sizeof(int), slice_data_address);
   }
 
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
   {
     Tensor src_tensor;
     src_tensor.mutable_data<double>(make_ddim({6, 9}), GPUPlace());
@@ -218,7 +218,7 @@ TEST(Tensor, CopyFrom) {
       EXPECT_EQ(dst_ptr[i], slice_ptr[i]);
     }
   }
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
   {
     Tensor src_tensor;
     Tensor gpu_tensor;

paddle/function/BlockExpandOp.cpp

@@ -194,7 +194,7 @@ public:
 
 REGISTER_TYPED_FUNC(BlockExpand, CPU, BlockExpandForward);
 REGISTER_TYPED_FUNC(BlockExpandGrad, CPU, BlockExpandBackward);
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 REGISTER_TYPED_FUNC(BlockExpand, GPU, BlockExpandForward);
 REGISTER_TYPED_FUNC(BlockExpandGrad, GPU, BlockExpandBackward);
 #endif

paddle/function/ContextProjectionOp.cpp

@@ -395,7 +395,7 @@ REGISTER_TYPED_FUNC(ContextProjectionForward,
 REGISTER_TYPED_FUNC(ContextProjectionBackward,
                     CPU,
                     ContextProjectionBackwardFunc);
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 REGISTER_TYPED_FUNC(ContextProjectionForward,
                     GPU,
                     ContextProjectionForwardFunc);

paddle/function/CosSimOp.cpp

@@ -233,7 +233,7 @@ private:
 
 REGISTER_TYPED_FUNC(CosSimForward, CPU, CosSimForwardFunc);
 REGISTER_TYPED_FUNC(CosSimBackward, CPU, CosSimBackwardFunc);
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 REGISTER_TYPED_FUNC(CosSimForward, GPU, CosSimForwardFunc);
 REGISTER_TYPED_FUNC(CosSimBackward, GPU, CosSimBackwardFunc);
 #endif

paddle/function/CropOp.cpp

@@ -169,7 +169,7 @@ private:
 
 REGISTER_TYPED_FUNC(Crop, CPU, CropFunc);
 REGISTER_TYPED_FUNC(CropGrad, CPU, CropGradFunc);
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 REGISTER_TYPED_FUNC(Crop, GPU, CropFunc);
 REGISTER_TYPED_FUNC(CropGrad, GPU, CropGradFunc);
 #endif

paddle/function/CrossMapNormalOp.cpp

@@ -336,7 +336,7 @@ private:
 
 REGISTER_TYPED_FUNC(CrossMapNormal, CPU, CrossMapNormalFunc);
 REGISTER_TYPED_FUNC(CrossMapNormalGrad, CPU, CrossMapNormalGradFunc);
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 REGISTER_TYPED_FUNC(CrossMapNormal, GPU, CrossMapNormalFunc);
 REGISTER_TYPED_FUNC(CrossMapNormalGrad, GPU, CrossMapNormalGradFunc);
 #endif

paddle/function/DepthwiseConvOp.cpp

@@ -292,7 +292,7 @@ REGISTER_TYPED_FUNC(DepthwiseConvGradInput,
 REGISTER_TYPED_FUNC(DepthwiseConvGradFilter,
                     CPU,
                     DepthwiseConvGradFilterFunction);
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 REGISTER_TYPED_FUNC(DepthwiseConv, GPU, DepthwiseConvFunction);
 REGISTER_TYPED_FUNC(DepthwiseConvGradInput,
                     GPU,

paddle/function/DepthwiseConvOpTest.cpp

@@ -17,7 +17,7 @@ limitations under the License. */
 
 namespace paddle {
 
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 TEST(DepthwiseConv, Forward) {
   DepthwiseConvolution<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU>(
       "GemmConv-CPU", "DepthwiseConv-GPU", forward);

paddle/function/GemmConvOp.cpp

@@ -340,7 +340,7 @@ public:
 REGISTER_TYPED_FUNC(GemmConv, CPU, GemmConvFunction);
 REGISTER_TYPED_FUNC(GemmConvGradInput, CPU, GemmConvGradInputFunction);
 REGISTER_TYPED_FUNC(GemmConvGradFilter, CPU, GemmConvGradFilterFunction);
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 REGISTER_TYPED_FUNC(GemmConv, GPU, GemmConvFunction);
 REGISTER_TYPED_FUNC(GemmConvGradInput, GPU, GemmConvGradInputFunction);
 REGISTER_TYPED_FUNC(GemmConvGradFilter, GPU, GemmConvGradFilterFunction);

paddle/function/GemmConvOpTest.cpp

@@ -24,7 +24,7 @@ TEST(GemmConv, NaiveConv) {
       "NaiveConv-CPU", "GemmConv-CPU", forward);
 }
 
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 TEST(GemmConv, Forward) {
   Convolution<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU>(
       "GemmConv-CPU", "GemmConv-GPU", forward);

paddle/function/Im2ColTest.cpp

@@ -116,7 +116,7 @@ void TestIm2ColFunctor() {
 
 TEST(Im2ColFunctor, CPU) { TestIm2ColFunctor<DEVICE_TYPE_CPU, float>(); }
 
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 
 TEST(Im2ColFunctor, GPU) { TestIm2ColFunctor<DEVICE_TYPE_GPU, float>(); }
 

paddle/function/MulOp.cpp

@@ -341,7 +341,7 @@ private:
 };
 
 REGISTER_TYPED_FUNC(MulOp, CPU, MulFunc);
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 REGISTER_TYPED_FUNC(MulOp, GPU, MulFunc);
 #endif
 }  // namespace paddle

paddle/function/PadOp.cpp

@@ -207,7 +207,7 @@ private:
 
 REGISTER_TYPED_FUNC(Pad, CPU, PadFunc);
 REGISTER_TYPED_FUNC(PadGrad, CPU, PadGradFunc);
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 REGISTER_TYPED_FUNC(Pad, GPU, PadFunc);
 REGISTER_TYPED_FUNC(PadGrad, GPU, PadGradFunc);
 #endif

paddle/function/RowConvOp.cpp

@@ -217,7 +217,7 @@ public:
 
 REGISTER_TYPED_FUNC(RowConv, CPU, RowConvFunc);
 REGISTER_TYPED_FUNC(RowConvGrad, CPU, RowConvGradFunc);
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 REGISTER_TYPED_FUNC(RowConv, GPU, RowConvFunc);
 REGISTER_TYPED_FUNC(RowConvGrad, GPU, RowConvGradFunc);
 #endif

paddle/function/SwitchOp.cpp

@@ -132,7 +132,7 @@ public:
 
 REGISTER_TYPED_FUNC(NCHW2NHWC, CPU, NCHW2NHWCFunc);
 REGISTER_TYPED_FUNC(NHWC2NCHW, CPU, NHWC2NCHWFunc);
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 REGISTER_TYPED_FUNC(NCHW2NHWC, GPU, NCHW2NHWCFunc);
 REGISTER_TYPED_FUNC(NHWC2NCHW, GPU, NHWC2NCHWFunc);
 #endif

paddle/gserver/layers/BatchNormBaseLayer.cpp

@@ -16,7 +16,7 @@ limitations under the License. */
 #include "BatchNormalizationLayer.h"
 #include "Layer.h"
 #include "paddle/utils/Stat.h"
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 #include "CudnnBatchNormLayer.h"
 #endif
 

paddle/gserver/layers/BatchNormalizationLayer.cpp

@@ -13,7 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/utils/Stat.h"
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 #include "hl_batch_transpose.h"
 #endif
 #include "BatchNormalizationLayer.h"
@@ -90,7 +90,7 @@ void BatchNormalizationLayer::expandMat(const MatrixPtr& in, MatrixPtr& out) {
   size_t batchSize = in->getHeight();
   CHECK_EQ(out->getHeight(), batchSize * imgPixels_);
   if (useGpu_) {
-#ifndef PADDLE_WITH_GPU
+#ifndef PADDLE_WITH_CUDA
     LOG(FATAL) << "paddle is compiled only for cpu";
 #else
     batchTranspose(
@@ -127,7 +127,7 @@ void BatchNormalizationLayer::shrinkMat(const MatrixPtr& in, MatrixPtr& out) {
   }
   CHECK_EQ(in->getHeight(), static_cast<size_t>(batchSize * imgPixels_));
   if (useGpu_) {
-#ifndef PADDLE_WITH_GPU
+#ifndef PADDLE_WITH_CUDA
     LOG(FATAL) << "paddle is compiled only for cpu";
 #else
     batchTranspose(

paddle/gserver/layers/PoolLayer.cpp

@@ -15,7 +15,7 @@ limitations under the License. */
 #include "PoolLayer.h"
 #include "PoolProjectionLayer.h"
 #include "paddle/utils/Logging.h"
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 #include "CudnnPoolLayer.h"
 #endif
 namespace paddle {
@@ -53,7 +53,7 @@ Layer* PoolLayer::create(const LayerConfig& config) {
   const std::string& pool = config.inputs(0).pool_conf().pool_type();
   if (pool == "max-projection" || pool == "avg-projection") {
     return new PoolProjectionLayer(config);
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
   } else if (CudnnPoolLayer::typeCheck(pool)) {
     return new CudnnPoolLayer(config);
 #endif

paddle/gserver/tests/LayerGradUtil.cpp

@@ -674,7 +674,7 @@ void testLayerGradKernel(TestConfig testConf,
                          bool useGpu,
                          bool useWeight,
                          float epsilon) {
-#ifndef PADDLE_WITH_GPU
+#ifndef PADDLE_WITH_CUDA
   if (useGpu) return;
 #endif
   FLAGS_use_gpu = useGpu;

paddle/gserver/tests/test_BatchNorm.cpp

@@ -119,7 +119,7 @@ TEST(Layer, batchNorm) {
   CHECK_EQ(static_cast<int>(convLayer->getOutputValue()->getWidth()), 576);
 }
 
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 void batchNormInference(int n, int c, int h, int w) {
   MatrixPtr input = std::make_shared<GpuMatrix>(n, c * h * w);
   MatrixPtr cudnnOut = std::make_shared<GpuMatrix>(n, c * h * w);

paddle/gserver/tests/test_ConvUnify.cpp

@@ -117,7 +117,7 @@ MatrixPtr doOneConvTest(size_t imgSize,
 }
 
 TEST(Layer, convParaUnified) {
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
   MatrixPtr input, resultCpu, resultGpu;
 
   /// TEST1 for conv ///

paddle/gserver/tests/test_DetectionOutput.cpp

@@ -150,7 +150,7 @@ TEST(Layer, detectionOutputLayerFwd) {
                            useGpu,
                            result2);
 
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
   // GPU case 1.
   useGpu = true;
   inputLoc = Matrix::create(1, 16, false, useGpu);

paddle/gserver/tests/test_Evaluator.cpp

@@ -51,7 +51,7 @@ void testEvaluator(TestConfig testConf,
                    string testEvaluatorName,
                    size_t batchSize,
                    bool useGpu) {
-#ifndef PADDLE_WITH_GPU
+#ifndef PADDLE_WITH_CUDA
   if (useGpu) return;
 #endif
   FLAGS_use_gpu = useGpu;

paddle/gserver/tests/test_KmaxSeqScore.cpp

@@ -97,7 +97,7 @@ TEST(Layer, kmaxSeqScoreLayer) {
       Matrix::create(subSeqStartPosition.back(), 1, false, false);
 
   std::vector<bool> mode = {false};
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
   mode.push_back(true);
 #endif
 

paddle/gserver/tests/test_LayerGrad.cpp

@@ -12,7 +12,7 @@ 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. */
 
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 #include <cudnn.h>
 #endif
 #include <gtest/gtest.h>
@@ -258,7 +258,7 @@ void testProjectionConv(size_t groups, bool isDeconv) {
                      true);
 }
 
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 TEST(Projection, conv) {
   /// test ConvProjection
   testProjectionConv(1, false);
@@ -422,7 +422,7 @@ TEST(Layer, depthwiseConvLayer) {
   //  'depthwise_conv' is a sepecial case of 'exconv' whose
   //  groups size equals to the input channels size.
   testDepthwiseConvLayer("exconv", /* useGpu= */ false);
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
   testDepthwiseConvLayer("exconv", /* useGpu= */ true);
 #endif
 }
@@ -480,7 +480,7 @@ void testConvLayer(const string& type, bool trans, bool useGpu) {
 
 TEST(Layer, convLayer) {
   testConvLayer("exconv", /* trans= */ false, /* useGpu= */ false);
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
   testConvLayer("exconv", /* trans= */ false, /* useGpu= */ true);
   testConvLayer("cudnn_conv", /* trans= */ false, /* useGpu= */ true);
 #endif
@@ -525,7 +525,7 @@ TEST(Layer, convTransLayer) {
   for (auto useGpu : {false, true}) {
     testConvTransLayer("exconvt", /* trans= */ false, /* useGpu= */ useGpu);
   }
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
   testConvTransLayer("cudnn_convt", /* trans= */ false, /* useGpu= */ true);
 #endif
 }
@@ -638,7 +638,7 @@ TEST(Layer, SelectiveFullyConnectedLayer) {
                 /* trans= */ false,
                 /* useGup= */ false,
                 false);
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
   testLayerGrad(config,
                 "selective_fc",
                 100,
@@ -1210,7 +1210,7 @@ void testPoolLayer(const string& poolType, bool trans, bool useGpu) {
   testLayerGrad(config, "pool", 100, trans, useGpu);
 }
 
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 void testPoolLayer2(const string& poolType, bool trans, bool useGpu) {
   TestConfig config;
   config.inputDefs.push_back({INPUT_DATA, "layer_0", 3200, 0});
@@ -1236,7 +1236,7 @@ TEST(Layer, PoolLayer) {
   testPoolLayer("avg-projection", /* trans= */ false, /* useGpu= */ false);
   testPoolLayer("max-projection", /* trans= */ false, /* useGpu= */ false);
 
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
   testPoolLayer("avg-projection", /* trans= */ false, /* useGpu= */ true);
   testPoolLayer("max-projection", /* trans= */ false, /* useGpu= */ true);
   testPoolLayer("cudnn-max-pool", /* trans= */ false, /* useGpu= */ true);
@@ -1309,7 +1309,7 @@ void testPool3DLayer(const string& poolType, bool trans, bool useGpu) {
 TEST(Layer, Pool3DLayer) {
   testPool3DLayer("avg", /* trans= */ false, /* useGpu= */ false);
   testPool3DLayer("max", /* trans= */ false, /* useGpu= */ false);
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
   testPool3DLayer("avg", /* trans= */ false, /* useGpu= */ true);
   testPool3DLayer("max", /* trans= */ false, /* useGpu= */ true);
 #endif
@@ -1695,7 +1695,7 @@ void testBatchNormLayer(const string& type, bool trans, bool useGpu) {
 
 TEST(Layer, BatchNormalizationLayer) {
   testBatchNormLayer("batch_norm", false, false);
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
   testBatchNormLayer("batch_norm", false, true);
   if (hl_get_cudnn_lib_version() >= int(4000)) {
     testBatchNormLayer("cudnn_batch_norm", false, true);
@@ -1744,7 +1744,7 @@ void testBatchNorm3DLayer(const string& type, bool trans, bool useGpu) {
 
 TEST(Layer, testBatchNorm3DLayer) {
   testBatchNorm3DLayer("batch_norm", false, false);
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
   testBatchNorm3DLayer("batch_norm", false, true);
   if (hl_get_cudnn_lib_version() >= int(4000)) {
     testBatchNorm3DLayer("cudnn_batch_norm", false, true);
@@ -2262,7 +2262,7 @@ void test3DConvLayer(const string& type, bool trans, bool useGpu) {
 
 TEST(Layer, test3DConvLayer) {
   test3DConvLayer("conv3d", /* trans= */ false, /* useGpu= */ false);
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
   test3DConvLayer("conv3d", /* trans= */ false, /* useGpu= */ true);
 #endif
 }
@@ -2339,7 +2339,7 @@ void test3DDeConvLayer(const string& type, bool trans, bool useGpu) {
 
 TEST(Layer, test3DDeConvLayer) {
   test3DDeConvLayer("deconv3d", /* trans= */ false, /* useGpu= */ false);
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
   test3DDeConvLayer("deconv3d", /* trans= */ false, /* useGpu= */ true);
 #endif
 }

paddle/gserver/tests/test_NetworkCompare.cpp

@@ -243,7 +243,7 @@ TEST(Compare, concat_slice) {
   compareNetwork(config_file_a, config_file_b);
 }
 
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 TEST(Compare, img_pool) {
   std::string config_file_a = "./gserver/tests/img_pool_a.conf";
   std::string config_file_b = "./gserver/tests/img_pool_b.conf";

paddle/gserver/tests/test_PriorBox.cpp

@@ -151,7 +151,7 @@ TEST(Layer, priorBoxLayerFwd) {
                     useGpu,
                     result);
 
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
   // reset the input parameters
   variance[1] = 0.1;
   variance[3] = 0.2;

paddle/gserver/tests/test_ProtoDataProvider.cpp

@@ -485,7 +485,7 @@ TEST(ProtoDataProvider, test) {
               // Currently in async mode, useGpu is not supported
               continue;
             }
-#ifndef PADDLE_WITH_GPU
+#ifndef PADDLE_WITH_CUDA
             if (useGpu) {
               continue;
             }
@@ -525,7 +525,7 @@ TEST(ProtoDataProvider, constant_slots) {
       for (int numConstantSlots : {1, 2}) {
         for (int useGpu : numTwoArray) {
           for (int dataCompression : numTwoArray) {
-#ifndef PADDLE_WITH_GPU
+#ifndef PADDLE_WITH_CUDA
             if (useGpu) {
               continue;
             }
@@ -708,7 +708,7 @@ TEST(ProtoSequenceDataProvider, test) {
               // Currently in async mode, useGpu is not supported
               continue;
             }
-#ifndef PADDLE_WITH_GPU
+#ifndef PADDLE_WITH_CUDA
             if (useGpu) {
               continue;
             }

paddle/gserver/tests/test_PyDataProvider.cpp

@@ -37,7 +37,7 @@ TEST(PyDataProvider, py_fill_slots) {
   config.clear_files();
   std::string dataFile = "gserver/tests/pyDataProvider/pyDataProviderList";
   config.set_files(dataFile);
-#ifndef PADDLE_WITH_GPU
+#ifndef PADDLE_WITH_CUDA
   bool useGpu = false;
 #else
   bool useGpu = true;
@@ -71,7 +71,7 @@ TEST(PyDataProvider, py_fill_nest_slots) {
   std::string dataFile = "gserver/tests/pyDataProvider/pyDataProviderList";
   config.set_files(dataFile);
   EXPECT_EQ(config.IsInitialized(), true);
-#ifndef PADDLE_WITH_GPU
+#ifndef PADDLE_WITH_CUDA
   bool useGpu = false;
 #else
   bool useGpu = true;

paddle/gserver/tests/test_SelectiveFCLayer.cpp

@@ -321,7 +321,7 @@ TEST(Layer, SelectiveFcLayer_train_dense_mul) {
       "filelist=gserver/tests/SelectiveFcTest/dense_mul_list";
 
   for (auto useGpu : {false, true}) {
-#ifndef PADDLE_WITH_GPU
+#ifndef PADDLE_WITH_CUDA
     if (useGpu) {
       break;
     }
@@ -388,7 +388,7 @@ void testSelectiveFcLayerTrainSparseMul(const LayerConfig& config,
                           outMatSelfc->getWidth(),
                           outMatSelfc->getElementCnt()));
   cpuOutMatSelfc->copyFrom(*outMatSelfc, HPPL_STREAM_DEFAULT);
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
   if (useGpu) {
     hl_stream_synchronize(HPPL_STREAM_DEFAULT);
   }
@@ -418,7 +418,7 @@ void testSelectiveFcLayerTrainSparseMul(const LayerConfig& config,
   MatrixPtr cpuOutMatFc(
       new CpuMatrix(outMatFc->getHeight(), outMatFc->getWidth()));
   cpuOutMatFc->copyFrom(*outMatFc, HPPL_STREAM_DEFAULT);
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
   if (useGpu) {
     hl_stream_synchronize(HPPL_STREAM_DEFAULT);
   }
@@ -443,7 +443,7 @@ TEST(Layer, SelectiveFcLayer_train_sparse_mul) {
   selLayerConfig.set_size(fcLayerWidth);
 
   testSelectiveFcLayerTrainSparseMul(selLayerConfig, false);
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
   testSelectiveFcLayerTrainSparseMul(selLayerConfig, true);
 #endif
 }

paddle/gserver/tests/test_SeqSliceLayerGrad.cpp

@@ -195,7 +195,7 @@ TEST(Layer, SeqSliceLayer) {
   vector<vector<real>> ends;
 
   std::vector<bool> mode = {false};
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
   mode.push_back(true);
 #endif
   genSeqInfo(seqStartPos, subSeqStartPos);

paddle/gserver/tests/test_WarpCTCLayer.cpp

@@ -199,7 +199,7 @@ TEST(Layer, WarpCTCLayer) {
     for (auto batchSize : {1, 10, 32}) {
       for (auto normByTimes : {false, true}) {
         for (auto useGpu : {false, true}) {
-#ifndef PADDLE_WITH_GPU
+#ifndef PADDLE_WITH_CUDA
           if (useGpu) continue;
 #endif
           LOG(INFO) << "layerSize=" << layerSize << " batchSize=" << batchSize

paddle/math/Matrix.cpp

@@ -670,7 +670,7 @@ void GpuMatrix::leftMul(Matrix& a, real scaleAB, real scaleT) {
 }
 
 void GpuMatrix::selectRows(Matrix& table, IVector& ids) {
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
   CHECK(dynamic_cast<GpuMatrix*>(&table));
   CHECK(table.useGpu());
   CHECK(ids.useGpu());
@@ -694,7 +694,7 @@ void GpuMatrix::selectRows(Matrix& table, IVector& ids) {
 }
 
 void GpuMatrix::addToRows(Matrix& table, IVector& ids) {
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
   CHECK(dynamic_cast<GpuMatrix*>(&table));
   CHECK(table.useGpu());
   CHECK(ids.useGpu());
@@ -741,7 +741,7 @@ void GpuMatrix::rowMax(Matrix& max) {
 }
 
 void GpuMatrix::rowMax(IVector& maxIds, Matrix& maxVal) {
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
   CHECK(maxIds.useGpu() && maxVal.useGpu()) << "Matrix type are not equal";
   size_t numSamples = getHeight();
   size_t beam = maxVal.getWidth();

paddle/math/SparseMatrix.cpp

@@ -836,7 +836,7 @@ void GpuSparseMatrix::zeroMem() {
 }
 
 void GpuSparseMatrix::rowMax(IVector& maxIds, Matrix& maxVal) {
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
   CHECK(maxIds.useGpu() && maxVal.useGpu()) << "Matrix type are not equal";
   size_t numSamples = getHeight();
   size_t beam = maxVal.getWidth();

paddle/math/Vector.cpp

@@ -172,7 +172,7 @@ void GpuVectorT<T>::isEqualTo(const VectorT<T>& b, const T& value) {
 
 template <class T>
 void GpuVectorT<T>::selectFrom(const VectorT<T>& src, const VectorT<int>& ids) {
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
   hl_vector_select_from<T>(this->getData(),
                            this->getSize(),
                            src.getData(),
@@ -850,7 +850,7 @@ CpuGpuVectorT<T>::CpuGpuVectorT(CpuGpuVectorT<T>& src,
                                 size_t size)
     : sync_(nullptr) {
   CHECK_LE(offset + size, static_cast<size_t>(src.getSize()));
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
   SyncedFlag* flag = src.getSync();
   if (*flag == DATA_AT_CPU) {
     src.copyToGpu();  // will set synchronous data between CPU and GPU
@@ -861,7 +861,7 @@ CpuGpuVectorT<T>::CpuGpuVectorT(CpuGpuVectorT<T>& src,
   auto cMemHandle = (src.getVector(false))->getMemoryHandle();
   cpuVectorT_ = std::make_shared<CpuVectorT<T>>(
       size, std::dynamic_pointer_cast<CpuMemoryHandle>(cMemHandle), offset);
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
   auto gMemHandle = (src.getVector(true))->getMemoryHandle();
   gpuVectorT_ = std::make_shared<GpuVectorT<T>>(
       size, std::dynamic_pointer_cast<GpuMemoryHandle>(gMemHandle), offset);

paddle/math/tests/test_Allocator.cpp

@@ -68,7 +68,7 @@ void testPoolAllocator() {
 
 TEST(Allocator, Pool) {
   testPoolAllocator<CpuAllocator>();
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
   testPoolAllocator<GpuAllocator>();
 #endif
 }
@@ -92,7 +92,7 @@ TEST(MemoryHandle, Cpu) {
   EXPECT_EQ(ptr1, ptr2);
 }
 
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 TEST(MemoryHandle, Gpu) {
   int numGpu = hl_get_device_count();
 

paddle/math/tests/test_BaseMatrix.cpp

@@ -12,7 +12,7 @@ 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. */
 
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 /**
  * This test file use autotest::AutoCompare and cmpWithoutArg to compares the
  * implementation of CPU and GPU member function in

paddle/math/tests/test_CpuGpuVector.cpp

@@ -12,7 +12,7 @@ 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. */
 
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 
 #include <gtest/gtest.h>
 #include "paddle/math/Vector.h"

paddle/math/tests/test_ExecViaCpu.cpp

@@ -94,7 +94,7 @@ void testWrapper(F&& f) {
   }
 }
 
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 TEST(ExecViaCpu, test1) {
   testWrapper(f);
   testWrapper(&f);

paddle/math/tests/test_GpuProfiler.cpp

@@ -12,7 +12,7 @@ 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. */
 
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 
 #include <gtest/gtest.h>
 #include "paddle/math/Matrix.h"

paddle/math/tests/test_Matrix.cpp

@@ -12,7 +12,7 @@ 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. */
 
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 /**
  * This test file use autotest::AutoCompare and cmpWithArg to compares the
  * implementation of CPU and GPU member function in Matrix.cpp.

paddle/math/tests/test_SparseMatrix.cpp

@@ -47,7 +47,7 @@ struct MatrixPara {
   SparseFormat format;
 };
 
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 void test_sparse_matrix_mul(MatrixPara paraA,
                             MatrixPara paraB,
                             MatrixPara paraC) {
@@ -452,7 +452,7 @@ TEST(Matrix, SparseMatrixCSRFormatTrimFrom) {
   matB->trimFrom(*mat);
   checkSMatrixEqual2(matA, matB);
 
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
   GpuSparseMatrixPtr matC = std::make_shared<GpuSparseMatrix>(
       height, trimedWidth, height, FLOAT_VALUE, SPARSE_CSR, true);
   matC->trimFrom(*mat);
@@ -546,7 +546,7 @@ TEST(Matrix, SparseMatrixCSCFormatTrimFrom) {
   matB->trimFrom(*mat);
   checkSMatrixEqual2(matA, matB);
 
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
   GpuSparseMatrixPtr matC = std::make_shared<GpuSparseMatrix>(
       height, trimedWidth, height, FLOAT_VALUE, SPARSE_CSC, true);
   matC->trimFrom(*mat);

paddle/math/tests/test_TrainingAlgorithm.cpp

@@ -91,7 +91,7 @@ int VectorCheckErr(const VectorPtr& vector1, const VectorPtr& vector2) {
 typedef std::function<void(size_t size, bool useGpu)> testMatrixFunc;
 
 void testCase(testMatrixFunc matrixFunc) {
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
   for (auto useGpu : {false, true}) {
 #else
   for (auto useGpu : {false}) {

paddle/math/tests/test_batchTranspose.cpp

@@ -17,7 +17,7 @@ limitations under the License. */
 
 using namespace paddle;  // NOLINT
 
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 TEST(MatrixBatchTransTest, test_batch_matrix_transpose) {
   const int nx = 100;
   const int ny = 50;

paddle/math/tests/test_matrixCompare.cpp

@@ -12,7 +12,7 @@ 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. */
 
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 /// This unittest checks GpuMatrix/CpuMatrix get same result, so disable when
 /// only cpu version.
 

paddle/math/tests/test_perturbation.cpp

@@ -12,7 +12,7 @@ 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. */
 
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 
 #include <cuda_runtime.h>
 #include <gtest/gtest.h>

paddle/math/tests/test_sparseMatrixCompare.cpp

@@ -12,7 +12,7 @@ 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. */
 
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 /// This unittest checks GpuSparseMatrix/CpuSparseMatrix get same result,
 //  so disable when
 /// only cpu version.

paddle/memory/detail/buddy_allocator.cc

@@ -175,7 +175,7 @@ void* BuddyAllocator::SystemAlloc(size_t size) {
 }
 
 BuddyAllocator::PoolSet::iterator BuddyAllocator::RefillPool() {
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
   if (system_allocator_->UseGpu()) {
     if ((total_used_ + total_free_) == 0) {
       // Compute the maximum allocation size for the first allocation.

paddle/memory/detail/system_allocator.cc

@@ -62,7 +62,7 @@ void CPUAllocator::Free(void* p, size_t size, size_t index) {
 
 bool CPUAllocator::UseGpu() const { return false; }
 
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 
 void* GPUAllocator::Alloc(size_t& index, size_t size) {
   // CUDA documentation doesn't explain if cudaMalloc returns nullptr

paddle/memory/detail/system_allocator.h

@@ -40,7 +40,7 @@ class CPUAllocator : public SystemAllocator {
   virtual bool UseGpu() const;
 };
 
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 class GPUAllocator : public SystemAllocator {
  public:
   virtual void* Alloc(size_t& index, size_t size);

paddle/memory/detail/system_allocator_test.cc

@@ -56,7 +56,7 @@ TEST(CPUAllocator, LockMem) {
   TestAllocator(a, 0);
 }
 
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 TEST(GPUAllocator, Alloc) {
   paddle::memory::detail::GPUAllocator a;
   TestAllocator(a, 2048);

paddle/memory/memcpy.cc

@@ -26,7 +26,7 @@ void Copy<platform::CPUPlace, platform::CPUPlace>(platform::CPUPlace, void* dst,
   std::memcpy(dst, src, num);
 }
 
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 template <>
 void Copy<platform::CPUPlace, platform::GPUPlace>(platform::CPUPlace dst_place,
                                                   void* dst,

paddle/memory/memcpy.h

@@ -33,7 +33,7 @@ namespace memory {
 template <typename DstPlace, typename SrcPlace>
 void Copy(DstPlace, void* dst, SrcPlace, const void* src, size_t num);
 
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 
 /**
  * \brief   Copy memory from one place to another place.

paddle/memory/memory.cc

@@ -62,7 +62,7 @@ size_t Used<platform::CPUPlace>(platform::CPUPlace place) {
   return GetCPUBuddyAllocator()->Used();
 }
 
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 
 BuddyAllocator* GetGPUBuddyAllocator(int gpu_id) {
   using BuddyAllocVec = std::vector<BuddyAllocator*>;

paddle/memory/memory_test.cc

@@ -80,7 +80,7 @@ TEST(BuddyAllocator, CPUMultAlloc) {
   }
 }
 
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 
 size_t align(size_t size, paddle::platform::GPUPlace place) {
   size += sizeof(paddle::memory::detail::Metadata);

paddle/operators/detail/strided_memcpy.h

@@ -34,7 +34,7 @@ struct StridedMemcpyFunctor<T, 1> {
       auto& cpu_place = boost::get<platform::CPUPlace>(place);
       memory::Copy(cpu_place, dst, cpu_place, src, sizeof(T) * dst_dim.head);
     } else {
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
       auto& gpu_place = boost::get<platform::GPUPlace>(place);
       auto& cuda_ctx =
           reinterpret_cast<const platform::CUDADeviceContext&>(dev_ctx);

paddle/operators/math/im2col_test.cc

@@ -71,7 +71,7 @@ void testIm2col() {
     context =
         new paddle::platform::CPUDeviceContext(paddle::platform::CPUPlace());
   } else {
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
     context =
         new paddle::platform::CUDADeviceContext(paddle::platform::GPUPlace());
 #else
@@ -116,7 +116,7 @@ void testIm2col() {
 
 TEST(math, im2col) {
   testIm2col<paddle::platform::CPUPlace>();
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
   testIm2col<paddle::platform::GPUPlace>();
 #endif
 }

paddle/operators/math/math_function_test.cc

 #include "paddle/operators/math/math_function.h"
 #include "gtest/gtest.h"
 
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 TEST(math_function, notrans_mul_trans) {
   paddle::framework::Tensor input1;
   paddle::framework::Tensor input1_gpu;

paddle/operators/strided_memcpy_test.cc

@@ -72,7 +72,7 @@ TEST(StridedMemcpy, CPUConcat) {
   }
 }
 
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 TEST(StridedMemcpy, GPUCrop) {
   // clang-format off
   int src[] = {

paddle/platform/device_context.cc

@@ -35,7 +35,7 @@ Eigen::DefaultDevice* CPUDeviceContext::eigen_device() const {
 
 Place CPUDeviceContext::GetPlace() const { return CPUPlace(); }
 
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 
 template <>
 Eigen::GpuDevice*

paddle/platform/device_context.h

@@ -14,7 +14,7 @@ limitations under the License. */
 #include "paddle/platform/enforce.h"
 #include "paddle/platform/place.h"
 
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 #include "paddle/platform/dynload/cublas.h"
 #include "paddle/platform/dynload/cudnn.h"
 #include "paddle/platform/gpu_info.h"
@@ -61,7 +61,7 @@ class CPUDeviceContext : public DeviceContext {
   std::unique_ptr<Eigen::DefaultDevice> eigen_device_;
 };
 
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 template <>
 struct EigenDeviceConverter<platform::GPUPlace> {
   using EigenDeviceType = Eigen::GpuDevice;

paddle/platform/enforce.h

@@ -29,7 +29,7 @@ limitations under the License. */
 #include <cxxabi.h>  // for __cxa_demangle
 #endif
 
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 
 #include "paddle/platform/dynload/cublas.h"
 #include "paddle/platform/dynload/cudnn.h"
@@ -113,7 +113,7 @@ inline typename std::enable_if<sizeof...(Args) != 0, void>::type throw_on_error(
   }
 }
 
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 
 template <typename... Args>
 inline typename std::enable_if<sizeof...(Args) != 0, void>::type throw_on_error(

paddle/platform/gpu_info.h

@@ -14,7 +14,7 @@ limitations under the License. */
 
 #pragma once
 
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 
 #include <cuda_runtime.h>
 #include <stddef.h>

paddle/platform/variant.h

@@ -16,7 +16,7 @@
 
 #include <boost/config.hpp>
 
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 
 // Because boost's variadic templates has bug on nvcc, boost will disable
 // variadic template support when GPU enabled on nvcc.

paddle/pserver/test/SocketTest.cpp

@@ -215,7 +215,7 @@ int main(int argc, char** argv) {
 
   uint64_t dataSize = FLAGS_dim * sizeof(real);
 
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
   GpuVector gpuParam(FLAGS_dim);
   GpuVector gpuGrad(FLAGS_dim);
 #else

paddle/pserver/test/test_ProtoServer.cpp

@@ -99,7 +99,7 @@ TEST(ProtoServer, regular) {
 }
 
 TEST(ProtoServer, extended) {
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
   ProtoClient* client;
   if (FLAGS_rdma_tcp == "rdma")
     client = new ProtoClient(FLAGS_server_addr, FLAGS_port, F_RDMA);

paddle/pybind/pybind.cc

@@ -34,7 +34,7 @@ static size_t UniqueIntegerGenerator() {
 }
 
 bool IsCompileGPU() {
-#ifndef PADDLE_WITH_GPU
+#ifndef PADDLE_WITH_CUDA
   return false;
 #else
   return true;
@@ -78,7 +78,7 @@ PYBIND11_PLUGIN(core) {
       .def("set", PyCPUTensorSetFromArray<float>)
       .def("set", PyCPUTensorSetFromArray<int>)
       .def("set", PyCPUTensorSetFromArray<double>)
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
       .def("set", PyCUDATensorSetFromArray<float>)
       .def("set", PyCUDATensorSetFromArray<int>)
       .def("set", PyCUDATensorSetFromArray<double>)
@@ -96,7 +96,7 @@ PYBIND11_PLUGIN(core) {
       .def(
           "__init__",
           [](LoDTensor &instance, const std::vector<std::vector<size_t>> &lod) {
-#ifndef PADDLE_WITH_GPU
+#ifndef PADDLE_WITH_CUDA
             new (&instance) LoDTensor(lod);
 #else
              LoD new_lod;
@@ -107,7 +107,7 @@ PYBIND11_PLUGIN(core) {
           })
       .def("set_lod",
            [](LoDTensor &self, const std::vector<std::vector<size_t>> &lod) {
-#ifndef PADDLE_WITH_GPU
+#ifndef PADDLE_WITH_CUDA
              self.set_lod(lod);
 #else
              LoD new_lod;
@@ -117,7 +117,7 @@ PYBIND11_PLUGIN(core) {
 #endif
            })
       .def("lod", [](LoDTensor &self) -> std::vector<std::vector<size_t>> {
-#ifndef PADDLE_WITH_GPU
+#ifndef PADDLE_WITH_CUDA
         return self.lod();
 #else
            auto lod = self.lod();
@@ -203,7 +203,7 @@ All parameter, weight, gradient are variables in Paddle.
       .def_static("create",
                   [](paddle::platform::GPUPlace& place)
                       -> paddle::platform::DeviceContext* {
-#ifndef PADDLE_WITH_GPU
+#ifndef PADDLE_WITH_CUDA
                     PADDLE_THROW("GPUPlace is not supported in CPU device.");
 #else
                     return new paddle::platform::CUDADeviceContext(place);

paddle/pybind/tensor_py.h

@@ -106,7 +106,7 @@ void PyCPUTensorSetFromArray(
   std::memcpy(dst, array.data(), sizeof(T) * array.size());
 }
 
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 template <typename T>
 void PyCUDATensorSetFromArray(
     framework::Tensor &self,

paddle/trainer/MergeModel.cpp

@@ -29,7 +29,7 @@ int main(int argc, char** argv) {
   initMain(argc, argv);
   initPython(argc, argv);
   string confFile = TrainerConfigHelper::getConfigNameFromPath(FLAGS_model_dir);
-#ifndef PADDLE_WITH_GPU
+#ifndef PADDLE_WITH_CUDA
   FLAGS_use_gpu = false;
 #endif
   auto config = std::make_shared<TrainerConfigHelper>(confFile);

paddle/trainer/tests/test_Compare.cpp

@@ -146,7 +146,7 @@ void compareGradient(comData& comDataCpu, comData& comDataGpu) {
 }
 
 int main(int argc, char** argv) {
-#ifndef PADDLE_WITH_GPU
+#ifndef PADDLE_WITH_CUDA
   exit(0);
 #endif
   paddle::initMain(argc, argv);

paddle/trainer/tests/test_CompareSparse.cpp

@@ -174,7 +174,7 @@ TEST(compareSparse, multiGradientMachine) {
     FLAGS_local = local;
     FLAGS_ports_num_for_sparse = 5;
     for (bool useGpu : {false, true}) {
-#ifndef PADDLE_WITH_GPU
+#ifndef PADDLE_WITH_CUDA
       if (useGpu) continue;
 #endif
       FLAGS_parallel_nn = useGpu;
@@ -198,7 +198,7 @@ TEST(compareSparse, NeuralNetwork) {
     FLAGS_local = local;
     FLAGS_ports_num_for_sparse = 5;
     for (bool useGpu : {false, true}) {
-#ifndef PADDLE_WITH_GPU
+#ifndef PADDLE_WITH_CUDA
       if (useGpu) continue;
 #endif
       FLAGS_parallel_nn = useGpu;

paddle/trainer/tests/test_Trainer.cpp

@@ -51,7 +51,7 @@ void checkGradientTest(const string& configFile,
 
 TEST(checkGradient, cpu) { checkGradientTest(configFile1, false, false); }
 
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 TEST(checkGradient, gpu) { checkGradientTest(configFile1, true, false); }
 
 TEST(checkGradient, multiGpu) {
@@ -97,7 +97,7 @@ TEST(checkGradient, hsigmoid) { checkGradientTest(configFile2, false, false); }
 
 TEST(checkGradient, chunk) {
   checkGradientTest(configFile3, false, false);
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
   checkGradientTest(configFile3, true, true);
 #endif
 }

paddle/trainer/tests/test_TrainerOnePass.cpp

@@ -79,7 +79,7 @@ void trainerOnePassTest(const string& configFile,
 // 1. test trainer (cpu, gpu).
 TEST(trainerOnePass, cpu) { trainerOnePassTest(configFile1, false, false); }
 
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 TEST(trainerOnePass, gpu) { trainerOnePassTest(configFile1, true, false); }
 
 TEST(trainerOnePass, gpu2) { trainerOnePassTest(configFile1, true, false, 2); }
@@ -94,7 +94,7 @@ TEST(trainerOnePass, parallel) {
 #endif
 
 // 2. test average_window.
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 TEST(average_window, gpu) {
   trainerOnePassTest(configFile1, true, false, 4, 0.01);
 }
@@ -266,7 +266,7 @@ TEST(checkRemoteUpdater, cpuTrainerOldUpdater) {
   checkRemoteParameterUpdaterTest(configFile1, false, false, 1, true);
 }
 
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
 TEST(checkRemoteUpdater, gpuTrainer) {
   checkRemoteParameterUpdaterTest(configFile1, true, false);
 }

paddle/trainer/tests/test_recurrent_machine_generation.cpp

@@ -113,7 +113,7 @@ void testGeneration(const string& configFile,
 #ifndef PADDLE_TYPE_DOUBLE
 
 TEST(RecurrentGradientMachine, test_generation) {
-#ifndef PADDLE_WITH_GPU
+#ifndef PADDLE_WITH_CUDA
   const auto useGpuConfs = {false};
 #else
   const auto useGpuConfs = {true, false};

paddle/utils/Flags.cpp

@@ -14,7 +14,7 @@ limitations under the License. */
 
 #include "Flags.h"
 
-#ifndef PADDLE_WITH_GPU
+#ifndef PADDLE_WITH_CUDA
 DEFINE_bool(use_gpu, false, "Only support CPU training");
 #else
 DEFINE_bool(use_gpu, true, "Whether to use GPU for training");

paddle/utils/Util.h

@@ -218,7 +218,7 @@ protected:
  * *d2* is peer device to enable direct access to by the d1 device.
  */
 inline void enablePeerAccess(int d1, int d2) {
-#ifdef PADDLE_WITH_GPU
+#ifdef PADDLE_WITH_CUDA
   if (hl_device_can_access_peer(d1, d2)) {
     SetDevice dev(d1);
     hl_device_enable_peer_access(d2);

paddle/utils/Version.h

@@ -48,7 +48,7 @@ void printVersion(std::ostream& os);
  * @return return true if paddle compiled with GPU
  */
 constexpr bool isWithGpu() {
-#ifndef PADDLE_WITH_GPU
+#ifndef PADDLE_WITH_CUDA
   return false;
 #else
   return true;