【PaddlePaddle Hackathon 4】No.63 add fp16 and bf16 for eye and frame (#…

…51819)

* add fp16 and bf16 for eye and frame

* fix bug

* fix bug

* fix bug

* Update test_frame_op.py

fix code style

* fix bug

* fix bug

paddle/phi/kernels/gpu/eye_kernel.cu

@@ -13,7 +13,6 @@
 // limitations under the License.
 
 #include "paddle/phi/kernels/eye_kernel.h"
-
 #include "paddle/phi/backends/gpu/gpu_context.h"
 #include "paddle/phi/core/kernel_registry.h"
 #include "paddle/phi/kernels/impl/eye_kernel_impl.h"
@@ -26,4 +25,5 @@ PD_REGISTER_KERNEL(eye,
                    double,
                    int64_t,
                    int,
-                   phi::dtype::float16) {}
+                   phi::dtype::float16,
+                   phi::dtype::bfloat16) {}

paddle/phi/kernels/gpu/frame_grad_kernel.cu

@@ -13,7 +13,6 @@
 // limitations under the License.
 
 #include "paddle/phi/kernels/frame_grad_kernel.h"
-
 #include "paddle/phi/backends/gpu/gpu_context.h"
 #include "paddle/phi/common/complex.h"
 #include "paddle/phi/core/kernel_registry.h"
@@ -28,5 +27,6 @@ PD_REGISTER_KERNEL(frame_grad,
                    float,
                    double,
                    phi::dtype::float16,
+                   phi::dtype::bfloat16,
                    phi::dtype::complex<float>,
                    phi::dtype::complex<double>) {}

paddle/phi/kernels/gpu/frame_kernel.cu

@@ -28,5 +28,6 @@ PD_REGISTER_KERNEL(frame,
                    float,
                    double,
                    phi::dtype::float16,
+                   phi::dtype::bfloat16,
                    phi::dtype::complex<float>,
                    phi::dtype::complex<double>) {}

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

@@ -21,29 +21,40 @@
 
 import paddle
 from paddle import fluid
-from paddle.fluid import framework
+from paddle.fluid import core, framework
 from paddle.fluid.framework import Program, program_guard
 
 
 class TestEyeOp(OpTest):
     def setUp(self):
         '''
-        Test eye op with specified shape
+        Test eye op with default shape
         '''
         self.python_api = paddle.eye
         self.op_type = "eye"
+        self.init_dtype()
+        self.init_attrs()
 
         self.inputs = {}
         self.attrs = {
-            'num_rows': 219,
-            'num_columns': 319,
-            'dtype': framework.convert_np_dtype_to_dtype_(np.int32),
+            'num_rows': self.num_columns,
+            'num_columns': self.num_columns,
+            'dtype': framework.convert_np_dtype_to_dtype_(self.dtype),
+        }
+        self.outputs = {
+            'Out': np.eye(self.num_rows, self.num_columns, dtype=self.dtype)
         }
-        self.outputs = {'Out': np.eye(219, 319, dtype=np.int32)}
 
     def test_check_output(self):
         self.check_output()
 
+    def init_dtype(self):
+        self.dtype = np.int32
+
+    def init_attrs(self):
+        self.num_rows = 319
+        self.num_columns = 319
+
 
 class TestEyeOp1(OpTest):
     def setUp(self):
@@ -178,6 +189,35 @@ def test_error(self):
             paddle.eye(-1)
 
 
+class TestEyeFP16OP(TestEyeOp):
+    '''Test eye op with specified dtype'''
+
+    def init_dtype(self):
+        self.dtype = np.float16
+
+
+@unittest.skipIf(
+    not core.is_compiled_with_cuda()
+    or not core.is_bfloat16_supported(core.CUDAPlace(0)),
+    "core is not complied with CUDA and not support the bfloat16",
+)
+class TestEyeBF16OP(OpTest):
+    def setUp(self):
+        self.op_type = "eye"
+        self.dtype = np.uint16
+        self.python_api = paddle.eye
+        self.inputs = {}
+        self.attrs = {
+            'num_rows': 219,
+            'num_columns': 319,
+        }
+        self.outputs = {'Out': np.eye(219, 319)}
+
+    def test_check_output(self):
+        place = core.CUDAPlace(0)
+        self.check_output_with_place(place)
+
+
 if __name__ == "__main__":
     paddle.enable_static()
     unittest.main()

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

@@ -15,10 +15,11 @@
 import unittest
 
 import numpy as np
-from eager_op_test import OpTest
+from eager_op_test import OpTest, convert_float_to_uint16
 from numpy.lib.stride_tricks import as_strided
 
 import paddle
+from paddle.fluid import core
 
 
 def frame_from_librosa(x, frame_length, hop_length, axis=-1):
@@ -48,23 +49,28 @@ class TestFrameOp(OpTest):
     def setUp(self):
         self.op_type = "frame"
         self.python_api = paddle.signal.frame
-        self.shape, self.type, self.attrs = self.initTestCase()
-        self.inputs = {
-            'X': np.random.random(size=self.shape).astype(self.type),
-        }
+
+        self.init_dtype()
+        self.init_shape()
+        self.init_attrs()
+
+        self.inputs = {'X': np.random.random(size=self.shape)}
         self.outputs = {
             'Out': frame_from_librosa(x=self.inputs['X'], **self.attrs)
         }
 
-    def initTestCase(self):
-        input_shape = (150,)
-        input_type = 'float64'
-        attrs = {
+    def init_dtype(self):
+        self.dtype = 'float64'
+
+    def init_shape(self):
+        self.shape = (150,)
+
+    def init_attrs(self):
+        self.attrs = {
             'frame_length': 50,
             'hop_length': 15,
             'axis': -1,
         }
-        return input_shape, input_type, attrs
 
     def test_check_output(self):
         paddle.enable_static()
@@ -137,5 +143,50 @@ def initTestCase(self):
         return input_shape, input_type, attrs
 
 
+class TestFrameFP16OP(TestFrameOp):
+    def init_dtype(self):
+        self.dtype = np.float16
+
+
+@unittest.skipIf(
+    not core.is_compiled_with_cuda()
+    or not core.is_bfloat16_supported(core.CUDAPlace(0)),
+    "core is not complied with CUDA and not support the bfloat16",
+)
+class TestFrameBF16OP(OpTest):
+    def setUp(self):
+        self.op_type = "frame"
+        self.python_api = paddle.signal.frame
+        self.shape, self.dtype, self.attrs = self.initTestCase()
+        x = np.random.random(size=self.shape).astype(np.float32)
+        out = frame_from_librosa(x, **self.attrs).copy()
+        self.inputs = {
+            'X': convert_float_to_uint16(x),
+        }
+        self.outputs = {'Out': convert_float_to_uint16(out)}
+
+    def initTestCase(self):
+        input_shape = (150,)
+        input_dtype = np.uint16
+        attrs = {
+            'frame_length': 50,
+            'hop_length': 15,
+            'axis': -1,
+        }
+        return input_shape, input_dtype, attrs
+
+    def test_check_output(self):
+        paddle.enable_static()
+        place = core.CUDAPlace(0)
+        self.check_output_with_place(place)
+        paddle.disable_static()
+
+    def test_check_grad_normal(self):
+        paddle.enable_static()
+        place = core.CUDAPlace(0)
+        self.check_grad_with_place(place, ['X'], 'Out')
+        paddle.disable_static()
+
+
 if __name__ == '__main__':
     unittest.main()