From 75c080b1762b8b83f6c2bb7baf95478a049b45d4 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Wed, 9 Nov 2016 06:23:42 -0800 Subject: Added a test to validate memory transfers between host and sycl device --- unsupported/test/cxx11_tensor_sycl.cpp | 71 +++++++++++++++++++++++++++------- 1 file changed, 56 insertions(+), 15 deletions(-) (limited to 'unsupported/test/cxx11_tensor_sycl.cpp') diff --git a/unsupported/test/cxx11_tensor_sycl.cpp b/unsupported/test/cxx11_tensor_sycl.cpp index 6a9c33422..05fbf9e62 100644 --- a/unsupported/test/cxx11_tensor_sycl.cpp +++ b/unsupported/test/cxx11_tensor_sycl.cpp @@ -27,7 +27,46 @@ using Eigen::SyclDevice; using Eigen::Tensor; using Eigen::TensorMap; -void test_sycl_cpu(const Eigen::SyclDevice &sycl_device) { +void test_sycl_mem_transfers(const Eigen::SyclDevice &sycl_device) { + int sizeDim1 = 100; + int sizeDim2 = 100; + int sizeDim3 = 100; + array tensorRange = {{sizeDim1, sizeDim2, sizeDim3}}; + Tensor in1(tensorRange); + Tensor out1(tensorRange); + Tensor out2(tensorRange); + Tensor out3(tensorRange); + + in1 = in1.random(); + + float* gpu_data1 = static_cast(sycl_device.allocate(in1.size()*sizeof(float))); + float* gpu_data2 = static_cast(sycl_device.allocate(out1.size()*sizeof(float))); + //float* gpu_data = static_cast(sycl_device.allocate(out2.size()*sizeof(float))); + + TensorMap> gpu1(gpu_data1, tensorRange); + TensorMap> gpu2(gpu_data2, tensorRange); + //TensorMap> 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)); + 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(); + + for (int i = 0; i < in1.size(); ++i) { + VERIFY_IS_APPROX(out1(i), in1(i) * 3.14f); + VERIFY_IS_APPROX(out2(i), in1(i) * 3.14f); + VERIFY_IS_APPROX(out3(i), in1(i) * 2.7f); + } + + sycl_device.deallocate(gpu_data1); + sycl_device.deallocate(gpu_data2); +} + +void test_sycl_computations(const Eigen::SyclDevice &sycl_device) { int sizeDim1 = 100; int sizeDim2 = 100; @@ -41,10 +80,10 @@ void test_sycl_cpu(const Eigen::SyclDevice &sycl_device) { in2 = in2.random(); in3 = in3.random(); - float * gpu_in1_data = static_cast(sycl_device.allocate(in1.dimensions().TotalSize()*sizeof(float))); - float * gpu_in2_data = static_cast(sycl_device.allocate(in2.dimensions().TotalSize()*sizeof(float))); - float * gpu_in3_data = static_cast(sycl_device.allocate(in3.dimensions().TotalSize()*sizeof(float))); - float * gpu_out_data = static_cast(sycl_device.allocate(out.dimensions().TotalSize()*sizeof(float))); + float * gpu_in1_data = static_cast(sycl_device.allocate(in1.size()*sizeof(float))); + float * gpu_in2_data = static_cast(sycl_device.allocate(in2.size()*sizeof(float))); + float * gpu_in3_data = static_cast(sycl_device.allocate(in3.size()*sizeof(float))); + float * gpu_out_data = static_cast(sycl_device.allocate(out.size()*sizeof(float))); TensorMap> gpu_in1(gpu_in1_data, tensorRange); TensorMap> gpu_in2(gpu_in2_data, tensorRange); @@ -53,7 +92,7 @@ void test_sycl_cpu(const Eigen::SyclDevice &sycl_device) { /// a=1.2f gpu_in1.device(sycl_device) = gpu_in1.constant(1.2f); - sycl_device.memcpyDeviceToHost(in1.data(), gpu_in1_data ,(in1.dimensions().TotalSize())*sizeof(float)); + sycl_device.memcpyDeviceToHost(in1.data(), gpu_in1_data ,(in1.size())*sizeof(float)); for (int i = 0; i < sizeDim1; ++i) { for (int j = 0; j < sizeDim2; ++j) { for (int k = 0; k < sizeDim3; ++k) { @@ -65,7 +104,7 @@ void test_sycl_cpu(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.dimensions().TotalSize())*sizeof(float)); + sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data ,(out.size())*sizeof(float)); for (int i = 0; i < sizeDim1; ++i) { for (int j = 0; j < sizeDim2; ++j) { for (int k = 0; k < sizeDim3; ++k) { @@ -77,9 +116,9 @@ void test_sycl_cpu(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.dimensions().TotalSize())*sizeof(float)); + sycl_device.memcpyHostToDevice(gpu_in2_data, in2.data(),(in2.size())*sizeof(float)); gpu_out.device(sycl_device) = gpu_in1 * gpu_in2; - sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data,(out.dimensions().TotalSize())*sizeof(float)); + sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data,(out.size())*sizeof(float)); for (int i = 0; i < sizeDim1; ++i) { for (int j = 0; j < sizeDim2; ++j) { for (int k = 0; k < sizeDim3; ++k) { @@ -93,7 +132,7 @@ void test_sycl_cpu(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.dimensions().TotalSize())*sizeof(float)); + sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data,(out.size())*sizeof(float)); for (int i = 0; i < sizeDim1; ++i) { for (int j = 0; j < sizeDim2; ++j) { for (int k = 0; k < sizeDim3; ++k) { @@ -107,7 +146,7 @@ void test_sycl_cpu(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.dimensions().TotalSize())*sizeof(float)); + sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data,(out.size())*sizeof(float)); for (int i = 0; i < sizeDim1; ++i) { for (int j = 0; j < sizeDim2; ++j) { for (int k = 0; k < sizeDim3; ++k) { @@ -121,7 +160,7 @@ void test_sycl_cpu(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.dimensions().TotalSize())*sizeof(float)); + sycl_device.memcpyDeviceToHost(out.data(),gpu_out_data,(out.size())*sizeof(float)); for (int i = 0; i < sizeDim1; ++i) { for (int j = 0; j < sizeDim2; ++j) { for (int k = 0; k < sizeDim3; ++k) { @@ -134,9 +173,9 @@ void test_sycl_cpu(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.dimensions().TotalSize())*sizeof(float)); + sycl_device.memcpyHostToDevice(gpu_in3_data, in3.data(),(in3.size())*sizeof(float)); 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.dimensions().TotalSize())*sizeof(float)); + sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data,(out.size())*sizeof(float)); for (int i = 0; i < sizeDim1; ++i) { for (int j = 0; j < sizeDim2; ++j) { for (int k = 0; k < sizeDim3; ++k) { @@ -152,8 +191,10 @@ void test_sycl_cpu(const Eigen::SyclDevice &sycl_device) { sycl_device.deallocate(gpu_in3_data); sycl_device.deallocate(gpu_out_data); } + void test_cxx11_tensor_sycl() { cl::sycl::gpu_selector s; Eigen::SyclDevice sycl_device(s); - CALL_SUBTEST(test_sycl_cpu(sycl_device)); + CALL_SUBTEST(test_sycl_mem_transfers(sycl_device)); + CALL_SUBTEST(test_sycl_computations(sycl_device)); } -- cgit v1.2.3