Add aten::conj_physical

src/ATen/native/xpu/UnaryOps.cpp

@@ -5,6 +5,7 @@
 #include <ATen/xpu/XPUNativeFunctions.h>
 
 #include <ATen/native/xpu/sycl/AbsKernel.h>
+#include <ATen/native/xpu/sycl/UnaryComplexKernels.h>
 #include <ATen/native/xpu/sycl/UnaryFractionKernels.h>
 #include <ATen/native/xpu/sycl/UnaryGeometricAcosKernel.h>
 #include <ATen/native/xpu/sycl/UnaryGeometricAcoshKernel.h>
@@ -515,4 +516,32 @@ Tensor& XPUNativeFunctions::erfc_out(const Tensor& self, Tensor& out) {
   return out;
 }
 
+Tensor& XPUNativeFunctions::conj_physical_out(const Tensor& self, Tensor& out) {
+  auto iter = TensorIterator::unary_op(out, self);
+  native::xpu::conj_physical_kernel(iter);
+  return out;
+}
+
+Tensor XPUNativeFunctions::_conj_physical(const Tensor& self) {
+  if (self.is_conj()) {
+    return self.conj().clone();
+  }
+  Tensor out;
+  auto iter = TensorIterator::unary_op(out, self);
+  native::xpu::conj_physical_kernel(iter);
+  return iter.output();
+}
+
+Tensor XPUNativeFunctions::conj_physical(const Tensor& self) {
+  if (!self.is_complex())
+    return self;
+  return XPUNativeFunctions::_conj_physical(self);
+}
+
+Tensor& XPUNativeFunctions::conj_physical_(Tensor& self) {
+  if (!self.is_complex())
+    return self;
+  return XPUNativeFunctions::conj_physical_out(self, self);
+}
+
 } // namespace at

src/ATen/native/xpu/XPUFallback.template

@@ -185,7 +185,6 @@ TORCH_LIBRARY_IMPL(aten, XPU, m) {
     "cholesky",
     "cholesky_inverse",
     "_cholesky_solve_helper",
-    "conj_physical.out",
     "copysign.out",
     "cosh.out",
     "count_nonzero.dim_IntList",

src/ATen/native/xpu/sycl/UnaryComplexKernels.cpp

@@ -30,6 +30,32 @@ void conj_kernel(TensorIterator& iter) {
       }));
 }
 
+template <typename scalar_t>
+struct ConjPhysicalFunctor {
+  scalar_t operator()(scalar_t z) const {
+    return std::conj(z);
+  }
+};
+
+template <typename TYPE>
+struct ConjPhysicalFunctor<c10::complex<TYPE>> {
+  c10::complex<TYPE> operator()(c10::complex<TYPE> z) const {
+    return c10::complex<TYPE>(z.real(), -z.imag());
+  }
+};
+
+void conj_physical_kernel(TensorIterator& iter) {
+  AT_DISPATCH_SWITCH(
+      iter.common_dtype(),
+      "conj_xpu",
+      AT_DISPATCH_CASE_ALL_TYPES_AND3(kBool, kBFloat16, kHalf, [&] {
+        // Conj is a no-op for non-complex types
+        copy_kernel(iter);
+      }) AT_DISPATCH_CASE_COMPLEX_TYPES_AND(kComplexHalf, [&] {
+        gpu_kernel(iter, ConjPhysicalFunctor<scalar_t>());
+      }));
+}
+
 template <typename scalar_t>
 struct NegConjScalarFunc {
   scalar_t operator()(scalar_t src_val) const {

src/ATen/native/xpu/sycl/UnaryComplexKernels.h

@@ -6,6 +6,8 @@ namespace at::native::xpu {
 
 void conj_kernel(TensorIterator& iter);
 
+void conj_physical_kernel(TensorIterator& iter);
+
 void neg_conj_kernel(TensorIterator& iter);
 
 void neg_kernel(TensorIterator& iter);

test/xpu/run_test_with_skip.py

@@ -2207,9 +2207,8 @@ def launch_test(test_case, skip_list=None, exe_list=None):
     # torch.autograd.gradcheck.GradcheckError: Jacobian computed with forward mode mismatch for output 0 with respect to input 0,
     "test_fn_fwgrad_bwgrad_nn_functional_rrelu_xpu_float64",
     "test_forward_mode_AD_nn_functional_rrelu_xpu_float64",
-    # RuntimeError: DispatchStub: unsupported device typexpu
-    "test_inplace_forward_mode_AD_conj_physical_xpu_complex128",
-    # NotImplementedError: Could not run 'aten::_to_dense' with arguments from the 'SparseXPU' backend.
+    "test_forward_mode_AD_conj_physical_xpu_complex128",
+# NotImplementedError: Could not run 'aten::_to_dense' with arguments from the 'SparseXPU' backend.
     "test_fn_fwgrad_bwgrad_to_sparse_xpu_float64",
     "test_forward_mode_AD_to_sparse_xpu_float64",
 )
@@ -2745,8 +2744,8 @@ def launch_test(test_case, skip_list=None, exe_list=None):
     ### Error #7 in TestBwdGradientsXPU , totally 2 , NotImplementedError: Could not run 'aten::_sparse_coo_tensor_with_dims_and_tensors' with arguments from the 'SparseXPU' backend. This could be because the operator doesn't exist for this backend, or was omitted during the selective/custom build process (if using custom build). If you are a Facebook employee using PyTorch on mobile, please visit https://fburl.com/ptmfixes for possible resolutions. 'aten::_sparse_coo_tensor_with_dims_and_tensors' is only available for these backends: [XPU, Meta, SparseCPU, SparseMeta, BackendSelect, Python, FuncTorchDynamicLayerBackMode, Functionalize, Named, Conjugate, Negative, ZeroTensor, ADInplaceOrView, AutogradOther, AutogradCPU, AutogradCUDA, AutogradHIP, AutogradXLA, AutogradMPS, AutogradIPU, AutogradXPU, AutogradHPU, AutogradVE, AutogradLazy, AutogradMTIA, AutogradPrivateUse1, AutogradPrivateUse2, AutogradPrivateUse3, AutogradMeta, AutogradNestedTensor, Tracer, AutocastCPU, AutocastXPU, AutocastCUDA, FuncTorchBatched, BatchedNestedTensor, FuncTorchVmapMode, Batched, VmapMode, FuncTorchGradWrapper, PythonTLSSnapshot, FuncTorchDynamicLayerFrontMode, PreDispatch, PythonDispatcher].
     "test_fn_grad_to_sparse_xpu_float64",
     "test_fn_gradgrad_to_sparse_xpu_float64",
-    ### Error #8 in TestBwdGradientsXPU , totally 2 , RuntimeError: DispatchStub: unsupported device typexpu
-    "test_inplace_grad_conj_physical_xpu_complex128",
+    # torch.autograd.gradcheck.GradcheckError: Jacobian computed with forward mode mismatch for output 0 with respect to input 0,
+    "test_fn_grad_conj_physical_xpu_complex128",
     "test_inplace_gradgrad_conj_physical_xpu_complex128",
 )
 res += launch_test("test_ops_gradients_xpu.py", skip_list)

yaml/xpu_functions.yaml

@@ -500,3 +500,7 @@ supported:
   - randperm.generator_out
   - _amp_foreach_non_finite_check_and_unscale_
   - _amp_update_scale_
+  - _conj_physical
+  - conj_physical
+  - conj_physical.out
+  - conj_physical_