[EinsumOp] Make EinsumOp support bfloat16. (#43085)

* change einsum_v2 as default and add new flags: FLAG_einsum_opt=1|0

* make EInsumOP support bf16

* add unittest for BF16

* add condition for test_BF16

* fix bugs

* fix

paddle/phi/kernels/funcs/eigen/broadcast.cc

@@ -11,6 +11,7 @@ distributed under the License is distributed on an "AS IS" BASIS,
 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/phi/common/bfloat16.h"
 #include "paddle/phi/common/float16.h"
 #include "paddle/phi/kernels/funcs/eigen/eigen_function.h"
 
@@ -73,6 +74,7 @@ struct EigenBroadcastGrad<Eigen::DefaultDevice, T, Rank> {
   template struct FUNCTOR<Eigen::DefaultDevice, T, 6>
 INSTANTIATION(EigenBroadcast, bool);
 INSTANTIATION(EigenBroadcast, dtype::float16);
+INSTANTIATION(EigenBroadcast, dtype::bfloat16);
 INSTANTIATION(EigenBroadcast, float);
 INSTANTIATION(EigenBroadcast, double);
 INSTANTIATION(EigenBroadcast, int);

paddle/phi/kernels/funcs/eigen/broadcast.cu

@@ -11,6 +11,7 @@ distributed under the License is distributed on an "AS IS" BASIS,
 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/phi/common/bfloat16.h"
 #include "paddle/phi/common/float16.h"
 #include "paddle/phi/kernels/funcs/eigen/eigen_function.h"
 
@@ -73,6 +74,7 @@ struct EigenBroadcastGrad<Eigen::GpuDevice, T, Rank> {
   template struct FUNCTOR<Eigen::GpuDevice, T, 6>
 INSTANTIATION(EigenBroadcast, bool);
 INSTANTIATION(EigenBroadcast, dtype::float16);
+INSTANTIATION(EigenBroadcast, dtype::bfloat16);
 INSTANTIATION(EigenBroadcast, float);
 INSTANTIATION(EigenBroadcast, double);
 INSTANTIATION(EigenBroadcast, int);

paddle/phi/kernels/gpu/einsum_grad_kernel.cu

@@ -24,4 +24,5 @@ PD_REGISTER_KERNEL(einsum_grad,
                    phi::EinsumGradKernel,
                    float,
                    double,
-                   phi::dtype::float16) {}
+                   phi::dtype::float16,
+                   phi::dtype::bfloat16) {}

paddle/phi/kernels/gpu/tile_kernel.cu

@@ -27,4 +27,5 @@ PD_REGISTER_KERNEL(tile,
                    double,
                    int,
                    int64_t,
-                   phi::dtype::float16) {}
+                   phi::dtype::float16,
+                   phi::dtype::bfloat16) {}

paddle/phi/kernels/impl/einsum_grad_impl.h

@@ -197,6 +197,7 @@ void EinsumGradKernel(const Context& dev_ctx,
     // release the cache tensor dTC to save memory right now. they are useless
     // now.
     cache.clear();
+    if (x_grad[0]) {
       *(x_grad[0]) = PerformTileAndReduction<T, Context>(dev_ctx,
                                                          labeltype,
                                                          labelshape,
@@ -204,6 +205,8 @@ void EinsumGradKernel(const Context& dev_ctx,
                                                          ellipsis_dims[0],
                                                          ops[0],
                                                          dA);
+    }
+    if (x_grad[1]) {
       *(x_grad[1]) = PerformTileAndReduction<T, Context>(dev_ctx,
                                                          labeltype,
                                                          labelshape,
@@ -212,5 +215,6 @@ void EinsumGradKernel(const Context& dev_ctx,
                                                          ops[1],
                                                          dB);
     }
+  }
 }
 }  // namespace phi

python/paddle/fluid/tests/unittests/test_einsum_v2.py

@@ -478,5 +478,23 @@ class TestStaticGraphShape(unittest.TestCase):
         self.assertEqual(C.shape, (-1, 384))
 
 
+class TestBF16(unittest.TestCase):
+    """
+    EinsumOp support bfloat16 type, add unittest here for the correctness.
+    """
+
+    def test_shape(self):
+        cuda_major = paddle.version.cuda().split('.')[0].strip()
+        if paddle.is_compiled_with_cuda() and int(cuda_major) >= 11:
+            """ MatmulKernel support bfloat16 only if cuda_major > 11.0.
+            """
+            A = paddle.to_tensor(np.array([1.0, 2.0])).astype(paddle.bfloat16)
+            A = A.cuda()
+            B = paddle.to_tensor(np.array([2.0, 3.0])).astype(paddle.bfloat16)
+            B = B.cuda()
+            C = paddle.einsum('i,i->', A, B)
+            self.assertEqual(C.item(), 8.0)
+
+
 if __name__ == "__main__":
     unittest.main()