Modifying TensorDeviceSycl.h to always create buffer of type uint8_t and convert them to the actual type at the execution on the device; adding the queue interface class to separate the lifespan of sycl queue and buffers,created for that queue, from Eigen::SyclDevice; modifying sycl tests to support the evaluation of the results for both row major and column major data layout on all different devices that are supported by Sycl{CPU; GPU; and Host}.

1 files changed, 56 insertions, 49 deletions

diff --git a/unsupported/test/cxx11_tensor_sycl.cpp b/unsupported/test/cxx11_tensor_sycl.cpp
index 05fbf9e62..bf115d652 100644
--- a/unsupported/test/cxx11_tensor_sycl.cpp
+++ b/unsupported/test/cxx11_tensor_sycl.cpp
@@ -26,35 +26,32 @@ using Eigen::array;
 using Eigen::SyclDevice;
 using Eigen::Tensor;
 using Eigen::TensorMap;
-
+template <typename DataType, int DataLayout>
 void test_sycl_mem_transfers(const Eigen::SyclDevice &sycl_device) {
   int sizeDim1 = 100;
-  int sizeDim2 = 100;
-  int sizeDim3 = 100;
+  int sizeDim2 = 10;
+  int sizeDim3 = 20;
   array<int, 3> tensorRange = {{sizeDim1, sizeDim2, sizeDim3}};
-  Tensor<float, 3> in1(tensorRange);
-  Tensor<float, 3> out1(tensorRange);
-  Tensor<float, 3> out2(tensorRange);
-  Tensor<float, 3> out3(tensorRange);
+  Tensor<DataType, 3, DataLayout> in1(tensorRange);
+  Tensor<DataType, 3, DataLayout> out1(tensorRange);
+  Tensor<DataType, 3, DataLayout> out2(tensorRange);
+  Tensor<DataType, 3, DataLayout> out3(tensorRange);
 
   in1 = in1.random();
 
-  float* gpu_data1  = static_cast<float*>(sycl_device.allocate(in1.size()*sizeof(float)));
-  float* gpu_data2  = static_cast<float*>(sycl_device.allocate(out1.size()*sizeof(float)));
-  //float* gpu_data =  static_cast<float*>(sycl_device.allocate(out2.size()*sizeof(float)));
+  DataType* gpu_data1  = static_cast<DataType*>(sycl_device.allocate(in1.size()*sizeof(DataType)));
+  DataType* gpu_data2  = static_cast<DataType*>(sycl_device.allocate(out1.size()*sizeof(DataType)));
+
+  TensorMap<Tensor<DataType, 3, DataLayout>> gpu1(gpu_data1, tensorRange);
+  TensorMap<Tensor<DataType, 3, DataLayout>> gpu2(gpu_data2, tensorRange);
 
-  TensorMap<Tensor<float, 3>> gpu1(gpu_data1, tensorRange);
-  TensorMap<Tensor<float, 3>> gpu2(gpu_data2, tensorRange);
-  //TensorMap<Tensor<float, 3>> gpu_out2(gpu_out2_data, tensorRange);
-  
-  sycl_device.memcpyHostToDevice(gpu_data1, in1.data(),(in1.size())*sizeof(float));
-  sycl_device.memcpyHostToDevice(gpu_data2, in1.data(),(in1.size())*sizeof(float));
+  sycl_device.memcpyHostToDevice(gpu_data1, in1.data(),(in1.size())*sizeof(DataType));
+  sycl_device.memcpyHostToDevice(gpu_data2, in1.data(),(in1.size())*sizeof(DataType));
   gpu1.device(sycl_device) = gpu1 * 3.14f;
   gpu2.device(sycl_device) = gpu2 * 2.7f;
-  sycl_device.memcpyDeviceToHost(out1.data(), gpu_data1,(out1.size())*sizeof(float));
-  sycl_device.memcpyDeviceToHost(out2.data(), gpu_data1,(out2.size())*sizeof(float));
-  sycl_device.memcpyDeviceToHost(out3.data(), gpu_data2,(out3.size())*sizeof(float));
-  //  sycl_device.Synchronize();
+  sycl_device.memcpyDeviceToHost(out1.data(), gpu_data1,(out1.size())*sizeof(DataType));
+  sycl_device.memcpyDeviceToHost(out2.data(), gpu_data1,(out2.size())*sizeof(DataType));
+  sycl_device.memcpyDeviceToHost(out3.data(), gpu_data2,(out3.size())*sizeof(DataType));
 
   for (int i = 0; i < in1.size(); ++i) {
     VERIFY_IS_APPROX(out1(i), in1(i) * 3.14f);
@@ -65,34 +62,34 @@ void test_sycl_mem_transfers(const Eigen::SyclDevice &sycl_device) {
   sycl_device.deallocate(gpu_data1);
   sycl_device.deallocate(gpu_data2);
 }
-
+template <typename DataType, int DataLayout>
 void test_sycl_computations(const Eigen::SyclDevice &sycl_device) {
 
   int sizeDim1 = 100;
-  int sizeDim2 = 100;
-  int sizeDim3 = 100;
+  int sizeDim2 = 10;
+  int sizeDim3 = 20;
   array<int, 3> tensorRange = {{sizeDim1, sizeDim2, sizeDim3}};
-  Tensor<float, 3> in1(tensorRange);
-  Tensor<float, 3> in2(tensorRange);
-  Tensor<float, 3> in3(tensorRange);
-  Tensor<float, 3> out(tensorRange);
+  Tensor<DataType, 3,DataLayout> in1(tensorRange);
+  Tensor<DataType, 3,DataLayout> in2(tensorRange);
+  Tensor<DataType, 3,DataLayout> in3(tensorRange);
+  Tensor<DataType, 3,DataLayout> out(tensorRange);
 
   in2 = in2.random();
   in3 = in3.random();
 
-  float * gpu_in1_data  = static_cast<float*>(sycl_device.allocate(in1.size()*sizeof(float)));
-  float * gpu_in2_data  = static_cast<float*>(sycl_device.allocate(in2.size()*sizeof(float)));
-  float * gpu_in3_data  = static_cast<float*>(sycl_device.allocate(in3.size()*sizeof(float)));
-  float * gpu_out_data =  static_cast<float*>(sycl_device.allocate(out.size()*sizeof(float)));
+  DataType * gpu_in1_data  = static_cast<DataType*>(sycl_device.allocate(in1.size()*sizeof(DataType)));
+  DataType * gpu_in2_data  = static_cast<DataType*>(sycl_device.allocate(in2.size()*sizeof(DataType)));
+  DataType * gpu_in3_data  = static_cast<DataType*>(sycl_device.allocate(in3.size()*sizeof(DataType)));
+  DataType * gpu_out_data =  static_cast<DataType*>(sycl_device.allocate(out.size()*sizeof(DataType)));
 
-  TensorMap<Tensor<float, 3>> gpu_in1(gpu_in1_data, tensorRange);
-  TensorMap<Tensor<float, 3>> gpu_in2(gpu_in2_data, tensorRange);
-  TensorMap<Tensor<float, 3>> gpu_in3(gpu_in3_data, tensorRange);
-  TensorMap<Tensor<float, 3>> gpu_out(gpu_out_data, tensorRange);
+  TensorMap<Tensor<DataType, 3, DataLayout>> gpu_in1(gpu_in1_data, tensorRange);
+  TensorMap<Tensor<DataType, 3, DataLayout>> gpu_in2(gpu_in2_data, tensorRange);
+  TensorMap<Tensor<DataType, 3, DataLayout>> gpu_in3(gpu_in3_data, tensorRange);
+  TensorMap<Tensor<DataType, 3, DataLayout>> gpu_out(gpu_out_data, tensorRange);
 
   /// a=1.2f
   gpu_in1.device(sycl_device) = gpu_in1.constant(1.2f);
-  sycl_device.memcpyDeviceToHost(in1.data(), gpu_in1_data ,(in1.size())*sizeof(float));
+  sycl_device.memcpyDeviceToHost(in1.data(), gpu_in1_data ,(in1.size())*sizeof(DataType));
   for (int i = 0; i < sizeDim1; ++i) {
     for (int j = 0; j < sizeDim2; ++j) {
       for (int k = 0; k < sizeDim3; ++k) {
@@ -104,7 +101,7 @@ void test_sycl_computations(const Eigen::SyclDevice &sycl_device) {
 
   /// a=b*1.2f
   gpu_out.device(sycl_device) = gpu_in1 * 1.2f;
-  sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data ,(out.size())*sizeof(float));
+  sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data ,(out.size())*sizeof(DataType));
   for (int i = 0; i < sizeDim1; ++i) {
     for (int j = 0; j < sizeDim2; ++j) {
       for (int k = 0; k < sizeDim3; ++k) {
@@ -116,9 +113,9 @@ void test_sycl_computations(const Eigen::SyclDevice &sycl_device) {
   printf("a=b*1.2f Test Passed\n");
 
   /// c=a*b
-  sycl_device.memcpyHostToDevice(gpu_in2_data, in2.data(),(in2.size())*sizeof(float));
+  sycl_device.memcpyHostToDevice(gpu_in2_data, in2.data(),(in2.size())*sizeof(DataType));
   gpu_out.device(sycl_device) = gpu_in1 * gpu_in2;
-  sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data,(out.size())*sizeof(float));
+  sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data,(out.size())*sizeof(DataType));
   for (int i = 0; i < sizeDim1; ++i) {
     for (int j = 0; j < sizeDim2; ++j) {
       for (int k = 0; k < sizeDim3; ++k) {
@@ -132,7 +129,7 @@ void test_sycl_computations(const Eigen::SyclDevice &sycl_device) {
 
   /// c=a+b
   gpu_out.device(sycl_device) = gpu_in1 + gpu_in2;
-  sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data,(out.size())*sizeof(float));
+  sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data,(out.size())*sizeof(DataType));
   for (int i = 0; i < sizeDim1; ++i) {
     for (int j = 0; j < sizeDim2; ++j) {
       for (int k = 0; k < sizeDim3; ++k) {
@@ -146,7 +143,7 @@ void test_sycl_computations(const Eigen::SyclDevice &sycl_device) {
 
   /// c=a*a
   gpu_out.device(sycl_device) = gpu_in1 * gpu_in1;
-  sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data,(out.size())*sizeof(float));
+  sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data,(out.size())*sizeof(DataType));
   for (int i = 0; i < sizeDim1; ++i) {
     for (int j = 0; j < sizeDim2; ++j) {
       for (int k = 0; k < sizeDim3; ++k) {
@@ -160,7 +157,7 @@ void test_sycl_computations(const Eigen::SyclDevice &sycl_device) {
 
   //a*3.14f + b*2.7f
   gpu_out.device(sycl_device) =  gpu_in1 * gpu_in1.constant(3.14f) + gpu_in2 * gpu_in2.constant(2.7f);
-  sycl_device.memcpyDeviceToHost(out.data(),gpu_out_data,(out.size())*sizeof(float));
+  sycl_device.memcpyDeviceToHost(out.data(),gpu_out_data,(out.size())*sizeof(DataType));
   for (int i = 0; i < sizeDim1; ++i) {
     for (int j = 0; j < sizeDim2; ++j) {
       for (int k = 0; k < sizeDim3; ++k) {
@@ -173,9 +170,9 @@ void test_sycl_computations(const Eigen::SyclDevice &sycl_device) {
   printf("a*3.14f + b*2.7f Test Passed\n");
 
   ///d= (a>0.5? b:c)
-  sycl_device.memcpyHostToDevice(gpu_in3_data, in3.data(),(in3.size())*sizeof(float));
+  sycl_device.memcpyHostToDevice(gpu_in3_data, in3.data(),(in3.size())*sizeof(DataType));
   gpu_out.device(sycl_device) =(gpu_in1 > gpu_in1.constant(0.5f)).select(gpu_in2, gpu_in3);
-  sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data,(out.size())*sizeof(float));
+  sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data,(out.size())*sizeof(DataType));
   for (int i = 0; i < sizeDim1; ++i) {
     for (int j = 0; j < sizeDim2; ++j) {
       for (int k = 0; k < sizeDim3; ++k) {
@@ -191,10 +188,20 @@ void test_sycl_computations(const Eigen::SyclDevice &sycl_device) {
   sycl_device.deallocate(gpu_in3_data);
   sycl_device.deallocate(gpu_out_data);
 }
-
+template<typename DataType, typename dev_Selector> void sycl_computing_test_per_device(dev_Selector s){
+  QueueInterface queueInterface(s);
+  auto sycl_device = Eigen::SyclDevice(&queueInterface);
+  test_sycl_mem_transfers<DataType, RowMajor>(sycl_device);
+  test_sycl_computations<DataType, RowMajor>(sycl_device);
+  test_sycl_mem_transfers<DataType, ColMajor>(sycl_device);
+  test_sycl_computations<DataType, ColMajor>(sycl_device);
+}
 void test_cxx11_tensor_sycl() {
-  cl::sycl::gpu_selector s;
-  Eigen::SyclDevice sycl_device(s);
-  CALL_SUBTEST(test_sycl_mem_transfers(sycl_device));
-  CALL_SUBTEST(test_sycl_computations(sycl_device));
+  printf("Test on GPU: OpenCL\n");
+  CALL_SUBTEST(sycl_computing_test_per_device<float>((cl::sycl::gpu_selector())));
+  printf("repeating the test on CPU: OpenCL\n");
+  CALL_SUBTEST(sycl_computing_test_per_device<float>((cl::sycl::cpu_selector())));
+  printf("repeating the test on CPU: HOST\n");
+  CALL_SUBTEST(sycl_computing_test_per_device<float>((cl::sycl::host_selector())));
+  printf("Test Passed******************\n" );
 }