Merge pull request #16 from EikanWang/master

Refine unit test case and write the shape meta info back to aten tensor when the tensor is reordered to public format

Author: EikanWang

scripts/cpu/gen-dense-cpu-ops.py

@@ -306,15 +306,17 @@ def is_out_func(fname):
                     if param_var == 'out' and is_out_func(fname):
                         code += '      TORCH_INTERNAL_ASSERT({}.is_contiguous());\n'.format(param_var)
                     else:
-                        # param_seq_str = '{}.is_contiguous() ? {} : {}.contiguous()'.format(param_var, param_var, param_var)
-                        None
+                        param_seq_str = '{}.is_contiguous() ? {} : {}.contiguous()'.format(param_var, param_var, param_var)
                 param_seq_str_vec.append(param_seq_str)
             code += '      if (dbl::chk::dnnl_support_the_tensors(dnnl_input_tensors))\n'
             code += '        return AtenIpexCPUDev::dil_{}({});\n'.format(fname, ', '.join(param_seq_str_vec))
 
         code += '    }\n'
 
         code += '  } catch (std::exception& e) {\n'
+        code += '#if defined(_DEBUG)\n'
+        code += '    TORCH_WARN(e.what());\n'
+        code += '#endif\n'
         code += '  }\n\n'
 
 

tests/cpu/test_lazy_reorder.py

@@ -368,12 +368,12 @@ def test_addbmm(self):
 
                 addbmm_cpu = torch.addbmm(res_cpu, b1_cpu, b2_cpu, beta=beta, alpha=alpha)
                 addbmm_dpcpp = torch.addbmm(res_dpcpp, b1_dpcpp, b2_dpcpp, beta=beta, alpha=alpha)
-                self.assertEqual(addbmm_cpu, addbmm_dpcpp)
+                self.assertEqual(addbmm_cpu, addbmm_dpcpp, 1e-4)
                 y_cpu = torch.randn(M, O, dtype=torch.float32)
                 y_dpcpp = y_cpu.to(device=device)
                 torch.addbmm(res_cpu, b1_cpu, b2_cpu, beta=beta, alpha=alpha, out=y_cpu)
                 torch.addbmm(res_dpcpp, b1_dpcpp, b2_dpcpp, beta=beta, alpha=alpha, out=y_dpcpp)
-                self.assertEqual(y_cpu, y_dpcpp)
+                self.assertEqual(y_cpu, y_dpcpp, 1e-4)
 
     def test_baddbmm(self):
         ipex.enable_auto_dnnl()
@@ -683,7 +683,6 @@ def test_batch_norm2d_backward(self):
 
         bn = torch.nn.BatchNorm2d(3)
         bn_dpcpp = copy.deepcopy(bn).to(device=device)
-
         y_cpu = bn(x_cpu).sum()
         y_dpcpp = bn_dpcpp(x_dpcpp).sum()
         y_cpu.backward()
@@ -756,17 +755,24 @@ def test_view(self):
 
         x_cpu = torch.randn(old_shape)
         x_dpcpp = x_cpu.to(device=device).clone()
-        print(x_dpcpp.size())
+        self.assertTrue(ipex.is_dil_tensor(x_dpcpp))
+        self.assertEqual(ipex.get_dil_tensor_sizes(x_dpcpp), [4, 16])
+        self.assertEqual(ipex.get_dil_tensor_strides(x_dpcpp), [16, 1])
 
         x_cpu_view = x_cpu.view(new_shape)
-        print(x_cpu_view.size())
+        self.assertEqual(x_cpu_view.size(), [1, 4, 4, 4])
+        self.assertEqual(x_cpu_view.stride(), [64, 16, 4, 1])
+
         x_dpcpp_view = x_dpcpp.view(new_shape)
-        print(x_dpcpp_view.size())
+        self.assertTrue(ipex.is_dil_tensor(x_dpcpp_view))
 
         y = torch.randn(new_shape)
         out_cpu = x_cpu_view * y
         # test if the shape of x_dpcpp_view is compatible with y
         out_dpcpp = x_dpcpp_view * y
+        self.assertTrue(ipex.is_dil_tensor(out_dpcpp))
+        self.assertEqual(ipex.get_dil_tensor_sizes(out_dpcpp), [1, 4, 4, 4])
+        self.assertEqual(ipex.get_dil_tensor_strides(out_dpcpp), [64, 16, 4, 1])
         self.assertEqual(out_cpu, out_dpcpp)
 
         # test if metadata of x_dpcpp has not been altered

torch_ipex/csrc/aten_ipex_bridge.cpp

@@ -11,6 +11,7 @@
 
 #include "ipex_tensor_impl.h"
 #include "ipex_sparse_tensor_impl.h"
+#include "cpu/dbl/Common.h"
 #include "cpu/ShadeDataContext.h"
 #include "cpu/bf16/Converter.h"
 #include "utils.h"
@@ -65,7 +66,6 @@ void reorderDilTensorToPublic(const at::Tensor& ipexTensor) {
   void *data_ctx = ipexTensor.unsafeGetTensorImpl()->storage().data_ptr().get_context();
   cpu::ShadeDataContext *shade_data_context = (cpu::ShadeDataContext*)data_ctx;
 #if defined(_DEBUG)
-  TORCH_WARN(ipexTensor.is_contiguous());
   TORCH_INTERNAL_ASSERT(! (shade_data_context->dil_tensor.is_empty()));
 #endif
   dil::tensor &dil_tensor = shade_data_context->dil_tensor;
@@ -101,7 +101,7 @@ void reorderDilTensorToPublic(const at::Tensor& ipexTensor) {
       ipexTensor.device().type());
 
     ipexTensor.unsafeGetTensorImpl()->storage().set_data_ptr(std::move(shade_data_ptr));
-    TORCH_INTERNAL_ASSERT(ipexTensor.is_contiguous());
+    cpu::dbl::comm::sync_shape_from_dil_to_aten(ipexTensor, pub_tensor);
   }
 }
 

torch_ipex/csrc/cpu/ShadeDataContext.h

@@ -90,9 +90,6 @@ struct ShadeDataContext {
     TORCH_INTERNAL_ASSERT((data_type == SHADE_DATA_TYPE::CPU_RAW) || (data_type == SHADE_DATA_TYPE::DIL));
 
     if (data_type == SHADE_DATA_TYPE::DIL) {
-#if defined(_DEBUG)
-      TORCH_WARN(tensor.is_contiguous());
-#endif
       auto raw_cpu_data = tensor.storage().data_ptr().get();
       if (raw_cpu_data == nullptr) {
         // the dnnl tensor does not share data with raw tensor data.
@@ -113,15 +110,11 @@ struct ShadeDataContext {
         //   C = A[4:7, :]
         // All these tensors share same buffer of Tensor A with different storge offsets and elements.
         // So the context modification will impact all these tensors.
-        if ((shade_data_context->dil_tensor.get_data_handle() == raw_cpu_data) &&
-            (shade_data_context->dil_tensor.get_nelems() == tensor.storage().numel()) &&
-            (shade_data_context->dil_tensor.get_data_type() == get_dil_data_type(tensor.scalar_type()))) {
-          //TODO: Do we need to check strides here?
+        if (check_tensor_own_whole_storage(tensor)) {
           TORCH_INTERNAL_ASSERT(shade_data_context->dil_tensor.get_size() == tensor.storage().capacity());
           return true;
         }
       }
-      TORCH_INTERNAL_ASSERT(false);
     }
 
     return false;
@@ -148,13 +141,9 @@ struct ShadeDataContext {
     TORCH_INTERNAL_ASSERT(tensor.has_storage());
     void *raw_context = tensor.storage().data_ptr().get_context();
     TORCH_INTERNAL_ASSERT(raw_context != nullptr);
-    if (isDilTensor(tensor)) {
-      ShadeDataContext *shade_data_context = (ShadeDataContext*)raw_context;
-      return shade_data_context->dil_tensor;
-    } else {
-      TORCH_INTERNAL_ASSERT(false);
-      return dil::tensor();
-    }
+    TORCH_INTERNAL_ASSERT(isDilTensor(tensor));
+    ShadeDataContext *shade_data_context = (ShadeDataContext*)raw_context;
+    return shade_data_context->dil_tensor;
   }
 
   /**

torch_ipex/csrc/cpu/dbl/Common.cpp

@@ -37,7 +37,7 @@ at::Tensor dil_tensor_to_dense(const at::Tensor& tensor) {
 dil::tensor try_gen_dil_tensor(const at::Tensor &input) {
   if (cpu::ShadeDataContext::isDilTensor(input)) {
     auto dil_tensor = cpu::ShadeDataContext::getDilTensor(input);
-    if (dil_tensor.is_public_format()) {
+    if ((!check_aten_dil_shape_info(input, dil_tensor)) && dil_tensor.is_public_format()) {
       dil_tensor.set_dims_and_strides(input.sizes().vec(), input.strides().vec());
     }
     return dil_tensor;

torch_ipex/csrc/init_python_bindings.cpp

@@ -12,6 +12,8 @@
 
 #include "aten_ipex_type.h"
 #include "auto_opt_config.h"
+#include "cpu/dil/dil.hpp"
+#include "cpu/ShadeDataContext.h"
 #include "cpu/ExtendOPs.h"
 #include "cpu/MlpOPs.h"
 
@@ -29,6 +31,28 @@ void setAutoDNNL(bool val) {
   AutoOptConfig::singleton().set_auto_dnnl(val);
 }
 
+/// **** Only for unit test ****
+bool isDilTensor(const at::Tensor &tensor) {
+  return cpu::ShadeDataContext::isDilTensor(tensor);
+}
+
+dil::dims getDilTensorSizes(const at::Tensor &tensor) {
+  if (isDilTensor(tensor)) {
+    auto dil_tensor = cpu::ShadeDataContext::getDilTensor(tensor);
+    return dil_tensor.get_dims();
+  }
+  return dil::dims();
+}
+
+dil::dims getDilTensorStrides(const at::Tensor &tensor) {
+  if (isDilTensor(tensor)) {
+    auto dil_tensor = cpu::ShadeDataContext::getDilTensor(tensor);
+    return dil_tensor.get_strides();
+  }
+  return dil::dims();
+}
+/// ****************************
+
 void InitIpexModuleBindings(py::module m) {
   m.def("_initialize_aten_bindings",
         []() { AtenIpexType::InitializeAtenBindings(); });
@@ -97,6 +121,10 @@ void InitIpexModuleBindings(py::module m) {
   m.def("mlp_create_handle", &AtenIpexTypeMLPExt::create_handle);
   m.def("mlp_set_relu_mask", &AtenIpexTypeMLPExt::set_relu_mask);
   m.def("mlp_release_handle", &AtenIpexTypeMLPExt::release_handle);
+
+  m.def("is_dil_tensor", &isDilTensor);
+  m.def("get_dil_tensor_sizes", &getDilTensorSizes);
+  m.def("get_dil_tensor_strides", &getDilTensorStrides);
 }
 
 }  // namespace

torch_ipex/csrc/utils.cpp

@@ -127,4 +127,13 @@ bool check_tensor_own_shade_context(const at::Tensor& tensor) {
   return (data_ptr != data_ctx) && (data_ctx != nullptr);
 }
 
+bool check_aten_dil_shape_info(const at::Tensor& ipex_tensor, const dil::tensor &dil_tensor) {
+  if (dil_tensor.is_public_format()) {
+    return ipex_tensor.sizes().vec() == dil_tensor.get_dims() &&
+          ipex_tensor.strides().vec() == dil_tensor.get_strides();
+  } else {
+    return ipex_tensor.sizes().vec() == dil_tensor.get_dims();
+  }
+}
+
 } // namespace torch_ipex

torch_ipex/csrc/utils.h

@@ -20,5 +20,6 @@ at::ScalarType get_at_data_type(dil::data_type);
 bool check_auto_dnnl();
 bool check_tensor_own_whole_storage(const at::Tensor& tensor);
 bool check_tensor_own_shade_context(const at::Tensor& tensor);
+bool check_aten_dil_shape_info(const at::Tensor& ipex_tensor, const dil::tensor &dil_tensor);
 
 } // namespace torch_ipex