Add matmul_v2 kernel in pten

paddle/fluid/framework/operator.cc

@@ -23,6 +23,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/data_type_transform.h"
 #include "paddle/fluid/framework/details/nan_inf_utils.h"
 #include "paddle/fluid/framework/op_call_stack.h"
+#include "paddle/fluid/framework/pten_utils.h"
 #include "paddle/fluid/framework/shape_inference.h"
 #include "paddle/fluid/framework/transfer_scope_cache.h"
 #include "paddle/fluid/framework/unused_var_check.h"

paddle/fluid/imperative/prepared_operator.cc

@@ -16,6 +16,7 @@
 
 #include "paddle/fluid/framework/data_type_transform.h"
 #include "paddle/fluid/framework/details/nan_inf_utils.h"
+#include "paddle/fluid/framework/pten_utils.h"
 #include "paddle/fluid/imperative/infer_shape_context.h"
 #include "paddle/pten/common/scalar.h"
 #include "paddle/utils/small_vector.h"

paddle/fluid/operators/matmul_v2_op.h

@@ -25,6 +25,11 @@ limitations under the License. */
 #include "paddle/fluid/operators/math/complex_functors.h"
 #include "paddle/fluid/operators/reduce_ops/reduce_sum_op.h"
 
+// only can include the headers in paddle/pten/api dirs
+#include "paddle/pten/api/include/core.h"
+#include "paddle/pten/api/include/linalg.h"
+#include "paddle/pten/hapi/lib/utils/tensor_utils.h"
+
 #if defined(__NVCC__) || defined(__HIPCC__)
 #include "paddle/fluid/operators/reduce_ops/cub_reduce.h"
 #endif
@@ -380,15 +385,17 @@ class MatMulV2Kernel : public framework::OpKernel<T> {
     auto* Out = ctx.Output<Tensor>("Out");
     bool trans_x = ctx.Attr<bool>("trans_x");
     bool trans_y = ctx.Attr<bool>("trans_y");
-    PADDLE_ENFORCE_NE(framework::product(X->dims()), 0,
-                      platform::errors::InvalidArgument(
-                          "The Input(X) dims size must not be equal 0,"
-                          " but reviced dims size is 0. "));
-    PADDLE_ENFORCE_NE(framework::product(Y->dims()), 0,
-                      platform::errors::InvalidArgument(
-                          "The Input(Y) dims size must not be equal 0,"
-                          " but reviced dims size is 0. "));
-    MatMulFunction<DeviceContext, T>(X, Y, Out, trans_x, trans_y, ctx);
+
+    auto& dev_ctx = ctx.device_context<DeviceContext>();
+    Out->mutable_data<T>(X->place());
+
+    auto pt_x = paddle::experimental::MakePtenDenseTensor(*X);
+    auto pt_y = paddle::experimental::MakePtenDenseTensor(*Y);
+    auto pt_out = paddle::experimental::MakePtenDenseTensor(*Out);
+
+    // call new kernel
+    pten::Matmul<T>(dev_ctx, *pt_x.get(), *pt_y.get(), trans_x, trans_y,
+                    pt_out.get());
   }
 };
 

paddle/pten/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

paddle/pten/hapi/lib/linalg.cc

@@ -25,7 +25,6 @@ limitations under the License. */
 #include "paddle/pten/core/kernel_context.h"
 #include "paddle/pten/hapi/lib/kernel_dispatch.h"
 #include "paddle/pten/hapi/lib/utils/allocator.h"
-#include "paddle/pten/infershape/binary.h"
 
 namespace paddle {
 namespace experimental {
@@ -65,5 +64,47 @@ Tensor dot(const Tensor& x, const Tensor& y) {
   return out;
 }
 
+Tensor matmul(const Tensor& x,
+              const Tensor& y,
+              bool transpose_x,
+              bool transpose_y) {
+  // 1. Get kernel signature and kernel
+  auto kernel_key_set = ParseKernelKeyByInputArgs(x, y);
+  auto kernel_key = kernel_key_set.GetHigestPriorityKernelKey();
+  auto kernel = pten::KernelFactory::Instance().SelectKernelOrThrowError(
+      "matmul_v2", kernel_key);
+
+  // 2. Get Device Context
+  auto* dev_ctx = GetDeviceContextByBackend(kernel_key.backend());
+  auto kernel_context = pten::KernelContext(*dev_ctx);
+
+  // 3. Auto data transform
+  auto dense_x = std::dynamic_pointer_cast<pten::DenseTensor>(x.impl());
+  auto dense_y = std::dynamic_pointer_cast<pten::DenseTensor>(y.impl());
+  kernel_context.EmplaceBackInput(dense_x);
+  kernel_context.EmplaceBackInput(dense_y);
+  kernel_context.EmplaceBackAttr(transpose_x);
+  kernel_context.EmplaceBackAttr(transpose_y);
+  // TODO(chenweihang): add transform impl
+
+  // 4. InferShape
+  auto out_meta = MatmulInferShape(
+      dense_x->meta(), dense_y->meta(), transpose_x, transpose_y);
+
+  // 5. Prepare outputs
+  const auto allocator = std::make_shared<DefaultAllocator>(
+      pten::TransToFluidPlace(kernel_key.backend()));
+  auto dense_out = std::make_shared<pten::DenseTensor>(allocator, out_meta);
+  kernel_context.EmplaceBackOutput(dense_out);
+
+  Tensor out;
+  out.set_impl(dense_out);
+
+  // 6. Call kernel
+  kernel(&kernel_context);
+
+  return out;
+}
+
 }  // namespace experimental
 }  // namespace paddle

paddle/pten/infershape/binary.cc

@@ -59,4 +59,74 @@ DenseTensorMeta DotInferShape(const DenseTensorMeta& x_meta,
   return return_meta;
 }
 
+DenseTensorMeta MatmulInferShape(const DenseTensorMeta& x_meta,
+                                 const DenseTensorMeta& y_meta,
+                                 bool trans_x,
+                                 bool trans_y) {
+  std::vector<int64_t> dims_x = paddle::framework::vectorize(x_meta.dims);
+  std::vector<int64_t> dims_y = paddle::framework::vectorize(y_meta.dims);
+  auto ndims_x = dims_x.size();
+  auto ndims_y = dims_y.size();
+  PADDLE_ENFORCE_GT(ndims_x,
+                    0,
+                    paddle::platform::errors::InvalidArgument(
+                        "The Input(x) dims size must be greater than 0,"
+                        " but reviced dims size is 0. "));
+  PADDLE_ENFORCE_GT(ndims_y,
+                    0,
+                    paddle::platform::errors::InvalidArgument(
+                        "The Input(y) dims size must be greater than 0,"
+                        " but reviced dims size is 0. "));
+
+  bool x_broadcasted = false, y_broadcasted = false;
+  if (ndims_x == 1) {
+    dims_x.insert(dims_x.begin(), 1);
+    ndims_x = 2;
+    x_broadcasted = true;
+  }
+
+  if (ndims_y == 1) {
+    dims_y.push_back(1);
+    ndims_y = 2;
+    y_broadcasted = true;
+  }
+
+  size_t M, N;
+  if (trans_x) {
+    M = dims_x[ndims_x - 1];
+  } else {
+    M = dims_x[ndims_x - 2];
+  }
+  if (trans_y) {
+    N = dims_y[ndims_y - 2];
+  } else {
+    N = dims_y[ndims_y - 1];
+  }
+
+  std::vector<int64_t> new_dims;
+  if (ndims_x > ndims_y) {
+    new_dims.assign(dims_x.begin(), dims_x.end() - 2);
+  } else if (ndims_x < ndims_y) {
+    new_dims.assign(dims_y.begin(), dims_y.end() - 2);
+  } else {
+    new_dims.reserve(ndims_x);
+    for (size_t i = 0; i < ndims_x - 2; ++i) {
+      new_dims.push_back(std::max(dims_x[i], dims_y[i]));
+    }
+  }
+  if (!x_broadcasted) {
+    new_dims.push_back(M);
+  }
+  if (!y_broadcasted) {
+    new_dims.push_back(N);
+  }
+  if (x_broadcasted && y_broadcasted) {
+    new_dims.push_back(1);
+  }
+
+  auto ddim_out = paddle::framework::make_ddim(new_dims);
+
+  return {x_meta.type, ddim_out, x_meta.layout};
+}
+
 }  // namespace pten

paddle/pten/infershape/binary.h

@@ -36,4 +36,9 @@ namespace pten {
 DenseTensorMeta DotInferShape(const DenseTensorMeta& x_meta,
                               const DenseTensorMeta& y_meta);
 
+DenseTensorMeta MatmulInferShape(const DenseTensorMeta& x_meta,
+                                 const DenseTensorMeta& y_meta,
+                                 bool trans_x,
+                                 bool trans_y);
+
 }  // namespace pten

paddle/pten/kernels/cpu/linalg.cc

@@ -21,6 +21,8 @@
 #include "paddle/fluid/operators/math/blas.h"
 #include "paddle/fluid/platform/complex.h"
 
+#include "paddle/pten/kernels/functions/math/matmul_func.h"
+
 namespace pten {
 
 template <typename T>
@@ -45,6 +47,27 @@ 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) {
+  PADDLE_ENFORCE_NE(paddle::framework::product(x.dims()),
+                    0,
+                    paddle::platform::errors::InvalidArgument(
+                        "The Input(X) dims size must not be equal 0,"
+                        " but reviced dims size is 0. "));
+  PADDLE_ENFORCE_NE(paddle::framework::product(y.dims()),
+                    0,
+                    paddle::platform::errors::InvalidArgument(
+                        "The Input(Y) dims size must not be equal 0,"
+                        " but reviced dims size is 0. "));
+  math::MatMulFunction<CPUContext, T>(
+      dev_ctx, x, y, out, transpose_x, transpose_y);
+}
+
 }  // namespace pten
 
 PT_REGISTER_MODULE(LinalgCPU);
@@ -62,3 +85,7 @@ PT_REGISTER_KERNEL("dot",
                    int64_t,
                    complex64,
                    complex128) {}
+
+PT_REGISTER_KERNEL(
+    "matmul_v2", CPU, ANY, pten::Matmul, float, double, complex64, complex128) {
+}

paddle/pten/kernels/cpu/linalg.h

@@ -30,7 +30,7 @@ void Dot(const CPUContext& dev_ctx,
          DenseTensor* out);
 
 template <typename T>
-void matmul(const CPUContext& dev_ctx,
+void Matmul(const CPUContext& dev_ctx,
             const DenseTensor& x,
             const DenseTensor& y,
             bool transpose_x,

paddle/pten/kernels/cuda/linalg.cu

@@ -16,6 +16,7 @@
 
 #include "paddle/pten/core/kernel_registry.h"
 #include "paddle/pten/kernels/functions/eigen/dot.h"
+#include "paddle/pten/kernels/functions/math/matmul_func.h"
 
 // See Note [ Why still include the fluid headers? ]
 #include "paddle/fluid/platform/complex.h"
@@ -30,10 +31,32 @@ void Dot(const CUDAContext& dev_ctx,
   eigen::Dot<CUDAContext, T>(dev_ctx, x, y, out);
 }
 
+template <typename T>
+void Matmul(const CUDAContext& dev_ctx,
+            const DenseTensor& x,
+            const DenseTensor& y,
+            bool transpose_x,
+            bool transpose_y,
+            DenseTensor* out) {
+  PADDLE_ENFORCE_NE(paddle::framework::product(x.dims()),
+                    0,
+                    paddle::platform::errors::InvalidArgument(
+                        "The Input(X) dims size must not be equal 0,"
+                        " but reviced dims size is 0. "));
+  PADDLE_ENFORCE_NE(paddle::framework::product(y.dims()),
+                    0,
+                    paddle::platform::errors::InvalidArgument(
+                        "The Input(Y) dims size must not be equal 0,"
+                        " but reviced dims size is 0. "));
+  math::MatMulFunction<CUDAContext, T>(
+      dev_ctx, x, y, out, transpose_x, transpose_y);
+}
+
 }  // namespace pten
 
 PT_REGISTER_MODULE(LinalgCUDA);
 
+using float16 = paddle::platform::float16;
 using complex64 = ::paddle::platform::complex<float>;
 using complex128 = ::paddle::platform::complex<double>;
 
@@ -47,3 +70,13 @@ PT_REGISTER_KERNEL("dot",
                    int64_t,
                    complex64,
                    complex128) {}
+
+PT_REGISTER_KERNEL("matmul_v2",
+                   CUDA,
+                   ANY,
+                   pten::Matmul,
+                   float,
+                   double,
+                   float16,
+                   complex64,
+                   complex128) {}

paddle/pten/kernels/cuda/linalg.h

@@ -32,6 +32,14 @@ void Dot(const CUDAContext& dev_ctx,
          const DenseTensor& y,
          DenseTensor* out);
 
+template <typename T>
+void Matmul(const CUDAContext& dev_ctx,
+            const DenseTensor& x,
+            const DenseTensor& y,
+            bool transpose_x,
+            bool transpose_y,
+            DenseTensor* out);
+
 }  // namespace pten
 
 #endif

paddle/pten/kernels/cuda/utils.h

@@ -14,6 +14,9 @@ limitations under the License. */
 
 #pragma once
 
+// CUDA and HIP use same api
+#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
+
 #include "paddle/pten/core/dense_tensor.h"
 #include "paddle/pten/core/kernel_registry.h"
 
@@ -26,3 +29,5 @@ using CUDAContext = paddle::platform::CUDADeviceContext;
 void Copy(const CUDAContext& dev_ctx, const DenseTensor& src, DenseTensor* dst);
 
 }  // namespace pten
+
+#endif