[Accuracy diff No.141] Fix accuracy diff for paddle.Tensor.mul API

paddle/phi/kernels/cpu/elementwise_multiply_kernel.cc

@@ -17,9 +17,9 @@
 #include "paddle/phi/common/bfloat16.h"
 #include "paddle/phi/common/complex.h"
 #include "paddle/phi/core/kernel_registry.h"
+#include "paddle/phi/kernels/cast_kernel.h"
 #include "paddle/phi/kernels/cpu/elementwise.h"
 #include "paddle/phi/kernels/impl/elementwise_kernel_impl.h"
-
 namespace phi {
 
 template <typename T, typename Context>
@@ -38,12 +38,59 @@ void MultiplyKernel(const Context& dev_ctx,
   } else {
     auto x_dims = x.dims();
     auto y_dims = y.dims();
-    if (x_dims.size() >= y_dims.size()) {
-      funcs::ElementwiseCompute<funcs::MultiplyFunctor<T>, T>(
-          dev_ctx, x, y, funcs::MultiplyFunctor<T>(), out, -1);
+    DenseTensor x_fp32 = phi::Cast<T, Context>(dev_ctx, x, DataType::FLOAT32);
+    DenseTensor y_fp32 = phi::Cast<T, Context>(dev_ctx, y, DataType::FLOAT32);
+    DataType final_out_dtype = out->dtype();
+    if (final_out_dtype == DataType::UNDEFINED) {
+      final_out_dtype = x.dtype();
+    }
+    if constexpr (std::is_same_v<T, phi::dtype::float16> ||
+                  std::is_same_v<T, phi::dtype::bfloat16>) {
+      if (final_out_dtype == DataType::FLOAT32) {
+        if (x_dims.size() >= y_dims.size()) {
+          funcs::ElementwiseCompute<funcs::MultiplyFunctor<T>, T>(
+              dev_ctx, x_fp32, y_fp32, funcs::MultiplyFunctor<T>(), out, -1);
+        } else {
+          funcs::ElementwiseCompute<funcs::InverseMultiplyFunctor<T>, T>(
+              dev_ctx,
+              x_fp32,
+              y_fp32,
+              funcs::InverseMultiplyFunctor<T>(),
+              out,
+              -1);
+        }
+      } else {
+        DenseTensor intermediate_result;
+        intermediate_result.set_meta(out->meta());
+        if (x_dims.size() >= y_dims.size()) {
+          funcs::ElementwiseCompute<funcs::MultiplyFunctor<T>, T>(
+              dev_ctx,
+              x_fp32,
+              y_fp32,
+              funcs::MultiplyFunctor<T>(),
+              &intermediate_result,
+              -1);
+        } else {
+          funcs::ElementwiseCompute<funcs::InverseMultiplyFunctor<T>, T>(
+              dev_ctx,
+              x_fp32,
+              y_fp32,
+              funcs::InverseMultiplyFunctor<T>(),
+              &intermediate_result,
+              -1);
+        }
+
+        phi::CastKernel<float, Context>(
+            dev_ctx, intermediate_result, final_out_dtype, out);
+      }
     } else {
-      funcs::ElementwiseCompute<funcs::InverseMultiplyFunctor<T>, T>(
-          dev_ctx, x, y, funcs::InverseMultiplyFunctor<T>(), out, -1);
+      if (x_dims.size() >= y_dims.size()) {
+        funcs::ElementwiseCompute<funcs::MultiplyFunctor<T>, T>(
+            dev_ctx, x, y, funcs::MultiplyFunctor<T>(), out, -1);
+      } else {
+        funcs::ElementwiseCompute<funcs::InverseMultiplyFunctor<T>, T>(
+            dev_ctx, x, y, funcs::InverseMultiplyFunctor<T>(), out, -1);
+      }
     }
   }
 }
@@ -67,4 +114,5 @@ PD_REGISTER_KERNEL(multiply,
                    bool,
                    complex64,
                    complex128,
+                   phi::dtype::float16,
                    phi::dtype::bfloat16) {}

test/legacy_test/test_math_op_patch_var_base.py

@@ -216,6 +216,15 @@ def test_mul(self):
             res = a * b
             np.testing.assert_array_equal(res.numpy(), a_np * b_np)
 
+    def test_mul_fp16(self):
+        a_np = np.random.random(self.shape).astype(np.float16)
+        b_np = np.random.random(self.shape).astype(np.float16)
+        with base.dygraph.guard():
+            a = paddle.to_tensor(a_np)
+            b = paddle.to_tensor(b_np)
+            res = a * b
+            np.testing.assert_array_equal(res.numpy(), a_np * b_np)
+
     def test_type_promotion_mul_F2_F4(self):
         a_np = np.random.random(self.shape).astype(np.float16)
         b_np = np.random.random(self.shape).astype(np.float32)