Move cpu, cuda and other device code into kernels (#15)

* fill_any_like kernel refactor * remove useless code of full_like c++ api * Support Scalar in Tensor Compute Library * add scalar in dygraph and static graph mode * keep the basic type for attr, instead of using scalar for all * merge the code * start refactor matmul * move cpu, cuda and other device modules into kernels * merge code * polish code in operator.cc
PaddlePaddle · Oct 15, 2021 · d3ab655 · d3ab655 · paddle-bot-old · Oct 15, 2021
1 parent e0322d5
commit d3ab655
Show file tree

Hide file tree

Showing 41 changed files with 78 additions and 58 deletions.
diff --git a/paddle/fluid/framework/operator.cc b/paddle/fluid/framework/operator.cc
@@ -1966,7 +1966,6 @@ pt::KernelContext OperatorWithKernel::ConstructPtKernelContext(
   }
 
   for (size_t i = 0; i < attr_defs.size(); ++i) {
-    paddle::any attr_item;
     if (attr_defs[i].type_index == std::type_index(typeid(pt::Scalar))) {
       // TODO(chenweihang): support other attrs
       // In principle, the attr required by the dynamic mode should be

diff --git a/paddle/tcmpt/CMakeLists.txt b/paddle/tcmpt/CMakeLists.txt
@@ -5,27 +5,8 @@ add_subdirectory(api)
 add_subdirectory(hapi)
 # tcmpt core components
 add_subdirectory(core)
-# tcmpt eigne functors, now paddle must compiled with eigen, but eigen just is
-# one backend dtype, we should support cropping it for lite
-add_subdirectory(eigen)
 # tcmpt kernels for diff device
-add_subdirectory(cpu)
-if(WITH_GPU OR WITH_ROCM)
-  # TODO(chenweihang): if hip can split from cuda impl, we should add hip dir
-  add_subdirectory(cuda)
-endif()
-# TODO(chenweihang): migrate MKLDNN Kernel in the second phase of the project
-if(WITH_MKLDNN)
-  add_subdirectory(mkldnn)
-endif()
-# TODO(chenweihang): migrate NPU Kernel in the second phase of the project
-if(WITH_ASCEND_CL)
-  add_subdirectory(npu)
-endif()
-# TODO(chenweihang): migrate XPU Kernel in the second phase of the project
-if(WITH_XPU)
-  add_subdirectory(xpu)
-endif()
+add_subdirectory(kernels)
 # tcmpt infershape
 add_subdirectory(infershape)
 # TODO(xingfeng): tcmpt inner module API designed by a high-performance team

diff --git a/paddle/tcmpt/api/include/creation.h b/paddle/tcmpt/api/include/creation.h
@@ -14,5 +14,5 @@
 
 #pragma once
 
-#include "paddle/tcmpt/cpu/creation.h"
-#include "paddle/tcmpt/cuda/creation.h"
+#include "paddle/tcmpt/kernels/cpu/creation.h"
+#include "paddle/tcmpt/kernels/cuda/creation.h"
diff --git a/paddle/tcmpt/api/include/linalg.h b/paddle/tcmpt/api/include/linalg.h
@@ -15,5 +15,5 @@
 #pragma once
 
 // See Note: [ How do we organize the kernel directory ]
-#include "paddle/tcmpt/cpu/linalg.h"
-#include "paddle/tcmpt/cuda/linalg.h"
+#include "paddle/tcmpt/kernels/cpu/linalg.h"
+#include "paddle/tcmpt/kernels/cuda/linalg.h"
diff --git a/paddle/tcmpt/api/include/manipulation.h b/paddle/tcmpt/api/include/manipulation.h
@@ -15,5 +15,5 @@
 #pragma once
 
 // See Note: [ How do we organize the kernel directory ]
-#include "paddle/tcmpt/cpu/manipulation.h"
-#include "paddle/tcmpt/cuda/manipulation.h"
+#include "paddle/tcmpt/kernels/cpu/manipulation.h"
+#include "paddle/tcmpt/kernels/cuda/manipulation.h"
diff --git a/paddle/tcmpt/api/include/math.h b/paddle/tcmpt/api/include/math.h
@@ -15,5 +15,5 @@ limitations under the License. */
 #pragma once
 
 // See Note: [ How do we organize the kernel directory ]
-#include "paddle/tcmpt/cpu/math.h"
-#include "paddle/tcmpt/cuda/math.h"
+#include "paddle/tcmpt/kernels/cpu/math.h"
+#include "paddle/tcmpt/kernels/cuda/math.h"
diff --git a/paddle/tcmpt/hapi/include/linalg.h b/paddle/tcmpt/hapi/include/linalg.h
@@ -21,5 +21,10 @@ namespace experimental {
 
 Tensor dot(const Tensor& x, const Tensor& y);
 
+Tensor matmul(const Tensor& x,
+              const Tensor& y,
+              bool transpose_x,
+              bool transpose_y);
+
 }  // namespace experimental
 }  // namespace paddle
diff --git a/paddle/tcmpt/kernels/CMakeLists.txt b/paddle/tcmpt/kernels/CMakeLists.txt
@@ -0,0 +1,18 @@
+# tcmpt kernels for diff device
+add_subdirectory(cpu)
+if(WITH_GPU OR WITH_ROCM)
+  # TODO(chenweihang): if hip can split from cuda impl, we should add hip dir
+  add_subdirectory(cuda)
+endif()
+# TODO(chenweihang): migrate MKLDNN Kernel in the second phase of the project
+if(WITH_MKLDNN)
+  add_subdirectory(mkldnn)
+endif()
+# TODO(chenweihang): migrate NPU Kernel in the second phase of the project
+if(WITH_ASCEND_CL)
+  add_subdirectory(npu)
+endif()
+# TODO(chenweihang): migrate XPU Kernel in the second phase of the project
+if(WITH_XPU)
+  add_subdirectory(xpu)
+endif()
diff --git a/paddle/tcmpt/eigen/CMakeLists.txt → ...tcmpt/kernels/common/eigen/CMakeLists.txt b/paddle/tcmpt/eigen/CMakeLists.txt → ...tcmpt/kernels/common/eigen/CMakeLists.txt
diff --git a/paddle/tcmpt/eigen/common.h → paddle/tcmpt/kernels/common/eigen/common.h b/paddle/tcmpt/eigen/common.h → paddle/tcmpt/kernels/common/eigen/common.h
diff --git a/paddle/tcmpt/eigen/dot.h → paddle/tcmpt/kernels/common/eigen/dot.h b/paddle/tcmpt/eigen/dot.h → paddle/tcmpt/kernels/common/eigen/dot.h
@@ -15,7 +15,7 @@ limitations under the License. */
 #pragma once
 
 #include "paddle/tcmpt/core/dense_tensor.h"
-#include "paddle/tcmpt/eigen/common.h"
+#include "paddle/tcmpt/kernels/common/eigen/common.h"
 
 // See Note [ Why still include the fluid headers? ]
 #include "paddle/fluid/operators/eigen/eigen_function.h"

diff --git a/paddle/tcmpt/eigen/fill.h → paddle/tcmpt/kernels/common/eigen/fill.h b/paddle/tcmpt/eigen/fill.h → paddle/tcmpt/kernels/common/eigen/fill.h
@@ -15,7 +15,7 @@ limitations under the License. */
 #pragma once
 
 #include "paddle/tcmpt/core/dense_tensor.h"
-#include "paddle/tcmpt/eigen/common.h"
+#include "paddle/tcmpt/kernels/common/eigen/common.h"
 
 // See Note [ Why still include the fluid headers? ]
 #include "paddle/fluid/operators/eigen/eigen_function.h"

diff --git a/paddle/tcmpt/eigen/mean.h → paddle/tcmpt/kernels/common/eigen/mean.h b/paddle/tcmpt/eigen/mean.h → paddle/tcmpt/kernels/common/eigen/mean.h
@@ -15,7 +15,7 @@ limitations under the License. */
 #pragma once
 
 #include "paddle/tcmpt/core/dense_tensor.h"
-#include "paddle/tcmpt/eigen/common.h"
+#include "paddle/tcmpt/kernels/common/eigen/common.h"
 
 // See Note [ Why still include the fluid headers? ]
 #include "paddle/fluid/operators/eigen/eigen_function.h"

diff --git a/paddle/tcmpt/eigen/scale.h → paddle/tcmpt/kernels/common/eigen/scale.h b/paddle/tcmpt/eigen/scale.h → paddle/tcmpt/kernels/common/eigen/scale.h
@@ -15,7 +15,7 @@ limitations under the License. */
 #pragma once
 
 #include "paddle/tcmpt/core/dense_tensor.h"
-#include "paddle/tcmpt/eigen/common.h"
+#include "paddle/tcmpt/kernels/common/eigen/common.h"
 
 // See Note [ Why still include the fluid headers? ]
 #include "paddle/fluid/operators/eigen/eigen_function.h"

diff --git a/paddle/tcmpt/eigen/sign.h → paddle/tcmpt/kernels/common/eigen/sign.h b/paddle/tcmpt/eigen/sign.h → paddle/tcmpt/kernels/common/eigen/sign.h
@@ -15,7 +15,7 @@ limitations under the License. */
 #pragma once
 
 #include "paddle/tcmpt/core/dense_tensor.h"
-#include "paddle/tcmpt/eigen/common.h"
+#include "paddle/tcmpt/kernels/common/eigen/common.h"
 
 // See Note [ Why still include the fluid headers? ]
 #include "paddle/fluid/operators/eigen/eigen_function.h"

diff --git a/paddle/tcmpt/cpu/CMakeLists.txt → paddle/tcmpt/kernels/cpu/CMakeLists.txt b/paddle/tcmpt/cpu/CMakeLists.txt → paddle/tcmpt/kernels/cpu/CMakeLists.txt
@@ -1,5 +1,5 @@
 if(WIN32)
-    set(CURRENT_BINARY_DIR ${PADDLE_BINARY_DIR}/paddle/tcmpt/cpu)
+    set(CURRENT_BINARY_DIR ${PADDLE_BINARY_DIR}/paddle/tcmpt/kernels/cpu)
     kernel_instantiate(creation.cc)
     kernel_instantiate(math.cc)
     kernel_instantiate(linalg.cc)

diff --git a/paddle/tcmpt/cpu/creation.cc → paddle/tcmpt/kernels/cpu/creation.cc b/paddle/tcmpt/cpu/creation.cc → paddle/tcmpt/kernels/cpu/creation.cc
@@ -12,10 +12,10 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include "paddle/tcmpt/cpu/creation.h"
+#include "paddle/tcmpt/kernels/cpu/creation.h"
 
 #include "paddle/tcmpt/core/kernel_registry.h"
-#include "paddle/tcmpt/eigen/fill.h"
+#include "paddle/tcmpt/kernels/common/eigen/fill.h"
 
 namespace pt {
 

diff --git a/paddle/tcmpt/cpu/creation.h → paddle/tcmpt/kernels/cpu/creation.h b/paddle/tcmpt/cpu/creation.h → paddle/tcmpt/kernels/cpu/creation.h
diff --git a/paddle/tcmpt/cpu/linalg.cc → paddle/tcmpt/kernels/cpu/linalg.cc b/paddle/tcmpt/cpu/linalg.cc → paddle/tcmpt/kernels/cpu/linalg.cc
@@ -12,12 +12,13 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include "paddle/tcmpt/cpu/linalg.h"
+#include "paddle/tcmpt/kernels/cpu/linalg.h"
 
 #include "paddle/tcmpt/core/kernel_registry.h"
 
 // See Note [ Why still include the fluid headers? ]
 #include "paddle/fluid/framework/eigen.h"
+#include "paddle/fluid/operators/math/blas.h"
 #include "paddle/fluid/platform/complex.h"
 
 namespace pt {
@@ -44,6 +45,14 @@ void Dot(const CPUContext& dev_ctx,
   }
 }
 
+template <typename T>
+void matmul(const CPUContext& dev_ctx,
+            const DenseTensor& x,
+            const DenseTensor& y,
+            bool transpose_x,
+            bool transpose_y,
+            DenseTensor* out) {}
+
 }  // namespace pt
 
 PT_REGISTER_MODULE(LinalgCPU);

diff --git a/paddle/tcmpt/cpu/linalg.h → paddle/tcmpt/kernels/cpu/linalg.h b/paddle/tcmpt/cpu/linalg.h → paddle/tcmpt/kernels/cpu/linalg.h
@@ -29,4 +29,12 @@ void Dot(const CPUContext& dev_ctx,
          const DenseTensor& y,
          DenseTensor* out);
 
+template <typename T>
+void matmul(const CPUContext& dev_ctx,
+            const DenseTensor& x,
+            const DenseTensor& y,
+            bool transpose_x,
+            bool transpose_y,
+            DenseTensor* out);
+
 }  // namespace pt
diff --git a/paddle/tcmpt/cpu/manipulation.cc → paddle/tcmpt/kernels/cpu/manipulation.cc b/paddle/tcmpt/cpu/manipulation.cc → paddle/tcmpt/kernels/cpu/manipulation.cc
@@ -12,9 +12,9 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include "paddle/tcmpt/cpu/manipulation.h"
-#include "paddle/tcmpt/cpu/utils.h"
+#include "paddle/tcmpt/kernels/cpu/manipulation.h"
 #include "paddle/tcmpt/infershape/unary.h"
+#include "paddle/tcmpt/kernels/cpu/utils.h"
 
 namespace pt {
 

diff --git a/paddle/tcmpt/cpu/manipulation.h → paddle/tcmpt/kernels/cpu/manipulation.h b/paddle/tcmpt/cpu/manipulation.h → paddle/tcmpt/kernels/cpu/manipulation.h
diff --git a/paddle/tcmpt/cpu/math.cc → paddle/tcmpt/kernels/cpu/math.cc b/paddle/tcmpt/cpu/math.cc → paddle/tcmpt/kernels/cpu/math.cc
@@ -12,11 +12,11 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include "paddle/tcmpt/cpu/math.h"
+#include "paddle/tcmpt/kernels/cpu/math.h"
 
-#include "paddle/tcmpt/eigen/mean.h"
-#include "paddle/tcmpt/eigen/scale.h"
-#include "paddle/tcmpt/eigen/sign.h"
+#include "paddle/tcmpt/kernels/common/eigen/mean.h"
+#include "paddle/tcmpt/kernels/common/eigen/scale.h"
+#include "paddle/tcmpt/kernels/common/eigen/sign.h"
 
 // See Note [ Why still include the fluid headers? ]
 #include "paddle/fluid/framework/eigen.h"

diff --git a/paddle/tcmpt/cpu/math.h → paddle/tcmpt/kernels/cpu/math.h b/paddle/tcmpt/cpu/math.h → paddle/tcmpt/kernels/cpu/math.h
diff --git a/paddle/tcmpt/cpu/utils.cc → paddle/tcmpt/kernels/cpu/utils.cc b/paddle/tcmpt/cpu/utils.cc → paddle/tcmpt/kernels/cpu/utils.cc
@@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#include "paddle/tcmpt/cpu/utils.h"
+#include "paddle/tcmpt/kernels/cpu/utils.h"
 #include "paddle/fluid/memory/memcpy.h"
 #include "paddle/tcmpt/core/convert_utils.h"
 #include "paddle/tcmpt/core/dtype.h"

diff --git a/paddle/tcmpt/cpu/utils.h → paddle/tcmpt/kernels/cpu/utils.h b/paddle/tcmpt/cpu/utils.h → paddle/tcmpt/kernels/cpu/utils.h
diff --git a/paddle/tcmpt/cuda/CMakeLists.txt → paddle/tcmpt/kernels/cuda/CMakeLists.txt b/paddle/tcmpt/cuda/CMakeLists.txt → paddle/tcmpt/kernels/cuda/CMakeLists.txt
@@ -1,5 +1,5 @@
 if(WIN32)
-    set(CURRENT_BINARY_DIR ${PADDLE_BINARY_DIR}/paddle/tcmpt/cuda)
+    set(CURRENT_BINARY_DIR ${PADDLE_BINARY_DIR}/paddle/tcmpt/kernels/cuda)
     kernel_instantiate(creation.cu)
     kernel_instantiate(math.cu)
     kernel_instantiate(linalg.cu)

diff --git a/paddle/tcmpt/cuda/creation.cu → paddle/tcmpt/kernels/cuda/creation.cu b/paddle/tcmpt/cuda/creation.cu → paddle/tcmpt/kernels/cuda/creation.cu
@@ -12,10 +12,10 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include "paddle/tcmpt/cuda/creation.h"
+#include "paddle/tcmpt/kernels/cuda/creation.h"
 
 #include "paddle/tcmpt/core/kernel_registry.h"
-#include "paddle/tcmpt/eigen/fill.h"
+#include "paddle/tcmpt/kernels/common/eigen/fill.h"
 
 namespace pt {
 

diff --git a/paddle/tcmpt/cuda/creation.h → paddle/tcmpt/kernels/cuda/creation.h b/paddle/tcmpt/cuda/creation.h → paddle/tcmpt/kernels/cuda/creation.h
diff --git a/paddle/tcmpt/cuda/linalg.cu → paddle/tcmpt/kernels/cuda/linalg.cu b/paddle/tcmpt/cuda/linalg.cu → paddle/tcmpt/kernels/cuda/linalg.cu
@@ -12,10 +12,10 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include "paddle/tcmpt/cuda/linalg.h"
+#include "paddle/tcmpt/kernels/cuda/linalg.h"
 
 #include "paddle/tcmpt/core/kernel_registry.h"
-#include "paddle/tcmpt/eigen/dot.h"
+#include "paddle/tcmpt/kernels/common/eigen/dot.h"
 
 // See Note [ Why still include the fluid headers? ]
 #include "paddle/fluid/platform/complex.h"

diff --git a/paddle/tcmpt/cuda/linalg.h → paddle/tcmpt/kernels/cuda/linalg.h b/paddle/tcmpt/cuda/linalg.h → paddle/tcmpt/kernels/cuda/linalg.h
diff --git a/paddle/tcmpt/cuda/manipulation.cu → paddle/tcmpt/kernels/cuda/manipulation.cu b/paddle/tcmpt/cuda/manipulation.cu → paddle/tcmpt/kernels/cuda/manipulation.cu
@@ -12,9 +12,9 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include "paddle/tcmpt/cuda/manipulation.h"
-#include "paddle/tcmpt/cuda/utils.h"
 #include "paddle/tcmpt/infershape/unary.h"
+#include "paddle/tcmpt/kernels/cuda/manipulation.h"
+#include "paddle/tcmpt/kernels/cuda/utils.h"
 
 namespace pt {
 

diff --git a/paddle/tcmpt/cuda/manipulation.h → paddle/tcmpt/kernels/cuda/manipulation.h b/paddle/tcmpt/cuda/manipulation.h → paddle/tcmpt/kernels/cuda/manipulation.h
diff --git a/paddle/tcmpt/cuda/math.cu → paddle/tcmpt/kernels/cuda/math.cu b/paddle/tcmpt/cuda/math.cu → paddle/tcmpt/kernels/cuda/math.cu
@@ -12,11 +12,11 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#include "paddle/tcmpt/cuda/math.h"
+#include "paddle/tcmpt/kernels/cuda/math.h"
 
-#include "paddle/tcmpt/eigen/mean.h"
-#include "paddle/tcmpt/eigen/scale.h"
-#include "paddle/tcmpt/eigen/sign.h"
+#include "paddle/tcmpt/kernels/common/eigen/mean.h"
+#include "paddle/tcmpt/kernels/common/eigen/scale.h"
+#include "paddle/tcmpt/kernels/common/eigen/sign.h"
 
 #ifdef __NVCC__
 #include "cub/cub.cuh"

diff --git a/paddle/tcmpt/cuda/math.h → paddle/tcmpt/kernels/cuda/math.h b/paddle/tcmpt/cuda/math.h → paddle/tcmpt/kernels/cuda/math.h
diff --git a/paddle/tcmpt/cuda/utils.cu → paddle/tcmpt/kernels/cuda/utils.cu b/paddle/tcmpt/cuda/utils.cu → paddle/tcmpt/kernels/cuda/utils.cu
@@ -16,7 +16,7 @@ limitations under the License. */
 #include "paddle/tcmpt/core/convert_utils.h"
 #include "paddle/tcmpt/core/dtype.h"
 #include "paddle/tcmpt/core/kernel_registry.h"
-#include "paddle/tcmpt/cuda/utils.h"
+#include "paddle/tcmpt/kernels/cuda/utils.h"
 
 namespace pt {
 

diff --git a/paddle/tcmpt/cuda/utils.h → paddle/tcmpt/kernels/cuda/utils.h b/paddle/tcmpt/cuda/utils.h → paddle/tcmpt/kernels/cuda/utils.h
diff --git a/paddle/tcmpt/mkldnn/CMakeLists.txt → paddle/tcmpt/kernels/mkldnn/CMakeLists.txt b/paddle/tcmpt/mkldnn/CMakeLists.txt → paddle/tcmpt/kernels/mkldnn/CMakeLists.txt
diff --git a/paddle/tcmpt/npu/CMakeLists.txt → paddle/tcmpt/kernels/npu/CMakeLists.txt b/paddle/tcmpt/npu/CMakeLists.txt → paddle/tcmpt/kernels/npu/CMakeLists.txt
diff --git a/paddle/tcmpt/xpu/CMakeLists.txt → paddle/tcmpt/kernels/xpu/CMakeLists.txt b/paddle/tcmpt/xpu/CMakeLists.txt → paddle/tcmpt/kernels/xpu/CMakeLists.txt
diff --git a/paddle/tcmpt/tests/test_copy_api.cc b/paddle/tcmpt/tests/test_copy_api.cc
@@ -16,7 +16,7 @@ limitations under the License. */
 #include <memory>
 
 #include "paddle/tcmpt/core/kernel_registry.h"
-#include "paddle/tcmpt/cpu/utils.h"
+#include "paddle/tcmpt/kernels/cpu/utils.h"
 
 #include "paddle/tcmpt/core/dense_tensor.h"