Add SYCL copy support to tensor (#6764)

* Add SYCL support to Tensor To method
* Check for non-contiguous sycl tensors (not supported)
---------
Co-authored-by: Sameer Sheorey <41028320+ssheorey@users.noreply.github.com>

cpp/open3d/core/CMakeLists.txt

@@ -56,6 +56,7 @@ target_sources(core PRIVATE
     kernel/ReductionCPU.cpp
     kernel/UnaryEW.cpp
     kernel/UnaryEWCPU.cpp
+    kernel/UnaryEWSYCL.cpp
     linalg/AddMM.cpp
     linalg/AddMMCPU.cpp
     linalg/Det.cpp

cpp/open3d/core/Device.h

@@ -99,6 +99,10 @@ class IsDevice {
     inline bool IsCUDA() const {
         return GetDevice().GetType() == Device::DeviceType::CUDA;
     }
+
+    inline bool IsSYCL() const {
+        return GetDevice().GetType() == Device::DeviceType::SYCL;
+    }
 };
 
 }  // namespace core

cpp/open3d/core/Tensor.cpp

@@ -748,7 +748,7 @@ Tensor Tensor::Contiguous() const {
 std::string Tensor::ToString(bool with_suffix,
                              const std::string& indent) const {
     std::ostringstream rc;
-    if (IsCUDA() || !IsContiguous()) {
+    if (IsCUDA() || IsSYCL() || !IsContiguous()) {
         Tensor host_contiguous_tensor = Contiguous().To(Device("CPU:0"));
         rc << host_contiguous_tensor.ToString(false, indent);
     } else {

cpp/open3d/core/kernel/UnaryEW.cpp

@@ -50,20 +50,28 @@ void Copy(const Tensor& src, Tensor& dst) {
                           src.GetShape(), dst.GetShape());
     }
 
-    // Disbatch to device
+    // Dispatch to device
     Device src_device = src.GetDevice();
     Device dst_device = dst.GetDevice();
-    if ((!src_device.IsCPU() && !src_device.IsCUDA()) ||
-        (!dst_device.IsCPU() && !dst_device.IsCUDA())) {
+    if ((!src_device.IsCPU() && !src_device.IsCUDA() && !src_device.IsSYCL()) ||
+        (!dst_device.IsCPU() && !dst_device.IsCUDA() && !dst_device.IsSYCL())) {
         utility::LogError("Copy: Unimplemented device");
     }
     if (src_device.IsCPU() && dst_device.IsCPU()) {
         CopyCPU(src, dst);
-    } else {
+    } else if ((src_device.IsCPU() || src_device.IsCUDA()) &&
+               (dst_device.IsCPU() || dst_device.IsCUDA())) {
 #ifdef BUILD_CUDA_MODULE
         CopyCUDA(src, dst);
 #else
         utility::LogError("Not compiled with CUDA, but CUDA device is used.");
+#endif
+    } else if ((src_device.IsCPU() || src_device.IsSYCL()) &&
+               (dst_device.IsCPU() || dst_device.IsSYCL())) {
+#ifdef BUILD_SYCL_MODULE
+        CopySYCL(src, dst);
+#else
+        utility::LogError("Not compiled with SYCL, but SYCL device is used.");
 #endif
     }
 }

cpp/open3d/core/kernel/UnaryEW.h

@@ -49,6 +49,10 @@ void CopyCPU(const Tensor& src, Tensor& dst);
 void CopyCUDA(const Tensor& src, Tensor& dst);
 #endif
 
+#ifdef BUILD_SYCL_MODULE
+void CopySYCL(const Tensor& src, Tensor& dst);
+#endif
+
 }  // namespace kernel
 }  // namespace core
 }  // namespace open3d

cpp/open3d/core/kernel/UnaryEWSYCL.cpp

@@ -0,0 +1,53 @@
+// ----------------------------------------------------------------------------
+// -                        Open3D: www.open3d.org                            -
+// ----------------------------------------------------------------------------
+// Copyright (c) 2018-2023 www.open3d.org
+// SPDX-License-Identifier: MIT
+// ----------------------------------------------------------------------------
+
+#include <cmath>
+#include <cstring>
+
+#include "open3d/core/Dtype.h"
+#include "open3d/core/MemoryManager.h"
+#include "open3d/core/SizeVector.h"
+#include "open3d/core/Tensor.h"
+#include "open3d/core/kernel/UnaryEW.h"
+#include "open3d/utility/Logging.h"
+
+namespace open3d {
+namespace core {
+namespace kernel {
+
+void CopySYCL(const Tensor& src, Tensor& dst) {
+    // It has been checked that
+    // - at least one of src or dst is SYCL device
+    SizeVector shape = src.GetShape();
+    Dtype src_dtype = src.GetDtype();
+    Dtype dst_dtype = dst.GetDtype();
+    Device dst_device = dst.GetDevice();
+    Device src_device = src.GetDevice();
+
+    if (src_dtype != dst_dtype) {
+        utility::LogError(
+                "CopySYCL: Dtype conversion from src to dst not implemented!");
+    }
+    if ((dst_device.IsSYCL() && !dst.IsContiguous()) ||
+        (src_device.IsSYCL() && !src.IsContiguous())) {
+        utility::LogError(
+                "CopySYCL: NonContiguous SYCL tensor Copy not implemented!");
+    }
+    Tensor src_conti = src.Contiguous();  // No op if already contiguous
+    if (dst.IsContiguous() && src.GetShape() == dst.GetShape() &&
+        src_dtype == dst_dtype) {
+        MemoryManager::Memcpy(dst.GetDataPtr(), dst_device,
+                              src_conti.GetDataPtr(), src_conti.GetDevice(),
+                              src_dtype.ByteSize() * shape.NumElements());
+    } else {
+        dst.CopyFrom(src_conti.To(dst_device));
+    }
+}
+
+}  // namespace kernel
+}  // namespace core
+}  // namespace open3d

cpp/tests/core/Tensor.cpp

@@ -30,12 +30,24 @@ INSTANTIATE_TEST_SUITE_P(Tensor,
                          TensorPermuteDevices,
                          testing::ValuesIn(PermuteDevices::TestCases()));
 
+class TensorPermuteDevicesWithSYCL : public PermuteDevices {};
+INSTANTIATE_TEST_SUITE_P(
+        Tensor,
+        TensorPermuteDevicesWithSYCL,
+        testing::ValuesIn(PermuteDevicesWithSYCL::TestCases()));
+
 class TensorPermuteDevicePairs : public PermuteDevicePairs {};
 INSTANTIATE_TEST_SUITE_P(
         Tensor,
         TensorPermuteDevicePairs,
         testing::ValuesIn(TensorPermuteDevicePairs::TestCases()));
 
+class TensorPermuteDevicePairsWithSYCL : public PermuteDevicePairsWithSYCL {};
+INSTANTIATE_TEST_SUITE_P(
+        Tensor,
+        TensorPermuteDevicePairsWithSYCL,
+        testing::ValuesIn(TensorPermuteDevicePairsWithSYCL::TestCases()));
+
 class TensorPermuteSizesDefaultStridesAndDevices
     : public testing::TestWithParam<
               std::tuple<std::pair<core::SizeVector, core::SizeVector>,
@@ -54,7 +66,7 @@ static constexpr const T &AsConst(T &t) noexcept {
     return t;
 }
 
-TEST_P(TensorPermuteDevices, Constructor) {
+TEST_P(TensorPermuteDevicesWithSYCL, Constructor) {
     core::Device device = GetParam();
     core::Dtype dtype = core::Float32;
 
@@ -71,7 +83,7 @@ TEST_P(TensorPermuteDevices, Constructor) {
     EXPECT_ANY_THROW(core::Tensor({-1, -1}, dtype, device));
 }
 
-TEST_P(TensorPermuteDevices, ConstructorBool) {
+TEST_P(TensorPermuteDevicesWithSYCL, ConstructorBool) {
     core::Device device = GetParam();
 
     core::SizeVector shape{2, 3};
@@ -105,7 +117,7 @@ TEST_P(TensorPermuteDevices, WithInitValue) {
     EXPECT_EQ(t.ToFlatVector<float>(), vals);
 }
 
-TEST_P(TensorPermuteDevices, WithInitList) {
+TEST_P(TensorPermuteDevicesWithSYCL, WithInitList) {
     core::Device device = GetParam();
 
     core::Tensor t;
@@ -187,23 +199,23 @@ TEST_P(TensorPermuteDevices, WithInitList) {
                  std::exception);
 }
 
-TEST_P(TensorPermuteDevices, WithInitValueBool) {
+TEST_P(TensorPermuteDevicesWithSYCL, WithInitValueBool) {
     core::Device device = GetParam();
 
     std::vector<bool> vals{true, false, true, true, false, false};
     core::Tensor t(vals, {2, 3}, core::Bool, device);
     EXPECT_EQ(t.ToFlatVector<bool>(), vals);
 }
 
-TEST_P(TensorPermuteDevices, WithInitValueTypeMismatch) {
+TEST_P(TensorPermuteDevicesWithSYCL, WithInitValueTypeMismatch) {
     core::Device device = GetParam();
 
     std::vector<int> vals{0, 1, 2, 3, 4, 5};
     EXPECT_THROW(core::Tensor(vals, {2, 3}, core::Float32, device),
                  std::runtime_error);
 }
 
-TEST_P(TensorPermuteDevices, WithInitValueSizeMismatch) {
+TEST_P(TensorPermuteDevicesWithSYCL, WithInitValueSizeMismatch) {
     core::Device device = GetParam();
 
     std::vector<float> vals{0, 1, 2, 3, 4};
@@ -298,7 +310,7 @@ TEST_P(TensorPermuteDevicePairs, IndexSetFillFancy) {
                                   0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1, 0}));
 }
 
-TEST_P(TensorPermuteDevicePairs, Copy) {
+TEST_P(TensorPermuteDevicePairsWithSYCL, Copy) {
     core::Device dst_device;
     core::Device src_device;
     std::tie(dst_device, src_device) = GetParam();
@@ -317,7 +329,7 @@ TEST_P(TensorPermuteDevicePairs, Copy) {
     EXPECT_EQ(dst_t.ToFlatVector<float>(), vals);
 }
 
-TEST_P(TensorPermuteDevicePairs, CopyBool) {
+TEST_P(TensorPermuteDevicePairsWithSYCL, CopyBool) {
     core::Device dst_device;
     core::Device src_device;
     std::tie(dst_device, src_device) = GetParam();
@@ -357,12 +369,15 @@ TEST_P(TensorPermuteDevicePairs, ToDevice) {
     core::Device src_device;
     std::tie(dst_device, src_device) = GetParam();
 
-    core::Tensor src_t = core::Tensor::Init<float>({0, 1, 2, 3}, src_device);
+    core::Tensor src_t =
+            core::Tensor::Init<float>({0.f, 1.f, 2.f, 3.f}, src_device);
     core::Tensor dst_t = src_t.To(dst_device);
     EXPECT_TRUE(dst_t.To(src_device).AllClose(src_t));
 
     EXPECT_ANY_THROW(src_t.To(core::Device("CPU:1")));
 
+    EXPECT_ANY_THROW(src_t.To(core::Device("SYCL:100")));
+
     EXPECT_ANY_THROW(src_t.To(core::Device("CUDA:-1")));
     EXPECT_ANY_THROW(src_t.To(core::Device("CUDA:100000")));
 }
@@ -529,7 +544,7 @@ TEST_P(TensorPermuteDevices, Flatten) {
     EXPECT_ANY_THROW(src_t.Flatten(2, 1));
 }
 
-TEST_P(TensorPermuteDevices, DefaultStrides) {
+TEST_P(TensorPermuteDevicesWithSYCL, DefaultStrides) {
     core::Device device = GetParam();
 
     core::Tensor t0({}, core::Float32, device);
@@ -663,7 +678,7 @@ TEST_P(TensorPermuteDevices, ItemAssign) {
     EXPECT_EQ(t[1][2][3].Item<float>(), 101);
 }
 
-TEST_P(TensorPermuteDevices, ToString) {
+TEST_P(TensorPermuteDevicesWithSYCL, ToString) {
     using ::testing::AnyOf;
     core::Device device = GetParam();
     core::Tensor t;
@@ -738,7 +753,7 @@ TEST_P(TensorPermuteDevices, ToString) {
  [True False False]])");
 }
 
-TEST_P(TensorPermuteDevicePairs, CopyContiguous) {
+TEST_P(TensorPermuteDevicePairsWithSYCL, CopyContiguous) {
     core::Device dst_device;
     core::Device src_device;
     std::tie(dst_device, src_device) = GetParam();

docker/Dockerfile.ci

@@ -94,7 +94,7 @@ RUN conda --version \
 # Activate open3d virtualenv
 # This works during docker build. It becomes the prefix of all RUN commands.
 # Ref: https://stackoverflow.com/a/60148365/1255535
-SHELL ["conda", "run", "-n", "open3d", "/bin/bash", "-c"]
+SHELL ["conda", "run", "-n", "open3d", "/bin/bash", "-o", "pipefail", "-c"]
 
 # Dependencies: cmake
 ENV PATH=${HOME}/${CMAKE_VERSION}/bin:${PATH}