【PaddlePaddle Hackathon 4】No.56 : add fp16 test and bf16 for poisson (#51662)

longranger2 · web-flow · commit 608a3f2842fd · 2023-07-31T10:38:13.000+08:00
* add fp16 and bf16 support for poisson

* add fp16 and bf16 support for searchsorted

* fix bug

* Update test_searchsorted_op.py

fix function name

* Update test_poisson_op.py

fix function name

* fix bug

* remove the searchorted

* Update test_poisson_op.py

* fix bug of TestPoissonBF16Op

* Update test_poisson_op.py

* Update test_poisson_op.py

* Update test_poisson_op.py

* fix bug of import

* fix bug
diff --git a/paddle/phi/kernels/gpu/poisson_grad_kernel.cu b/paddle/phi/kernels/gpu/poisson_grad_kernel.cu
@@ -15,5 +15,11 @@
 #include "paddle/phi/core/kernel_registry.h"
 #include "paddle/phi/kernels/impl/poisson_grad_kernel_impl.h"
 
-PD_REGISTER_KERNEL(
-    poisson_grad, GPU, ALL_LAYOUT, phi::PoissonGradKernel, float, double) {}
+PD_REGISTER_KERNEL(poisson_grad,
+                   GPU,
+                   ALL_LAYOUT,
+                   phi::PoissonGradKernel,
+                   float,
+                   double,
+                   phi::dtype::float16,
+                   phi::dtype::bfloat16) {}
diff --git a/paddle/phi/kernels/gpu/poisson_kernel.cu b/paddle/phi/kernels/gpu/poisson_kernel.cu
@@ -64,5 +64,11 @@ void PoissonKernel(const Context& ctx, const DenseTensor& x, DenseTensor* out) {
 
 }  // namespace phi
 
-PD_REGISTER_KERNEL(
-    poisson, GPU, ALL_LAYOUT, phi::PoissonKernel, float, double) {}
+PD_REGISTER_KERNEL(poisson,
+                   GPU,
+                   ALL_LAYOUT,
+                   phi::PoissonKernel,
+                   float,
+                   double,
+                   phi::dtype::float16,
+                   phi::dtype::bfloat16) {}
diff --git a/test/legacy_test/test_poisson_op.py b/test/legacy_test/test_poisson_op.py
@@ -16,9 +16,14 @@
 import unittest
 
 import numpy as np
-from eager_op_test import OpTest
+from eager_op_test import (
+    OpTest,
+    convert_float_to_uint16,
+    convert_uint16_to_float,
+)
 
 import paddle
+from paddle.fluid import core
 
 paddle.enable_static()
 paddle.seed(100)
@@ -42,17 +47,20 @@ class TestPoissonOp1(OpTest):
     def setUp(self):
         self.op_type = "poisson"
         self.python_api = paddle.tensor.poisson
+        self.init_dtype()
         self.config()
 
         self.attrs = {}
         self.inputs = {'X': np.full([2048, 1024], self.lam, dtype=self.dtype)}
         self.outputs = {'Out': np.ones([2048, 1024], dtype=self.dtype)}
 
+    def init_dtype(self):
+        self.dtype = "float64"
+
     def config(self):
         self.lam = 10
         self.a = 5
         self.b = 15
-        self.dtype = "float64"
 
     def verify_output(self, outs):
         hist, prob = output_hist(np.array(outs[0]), self.lam, self.a, self.b)
@@ -368,5 +376,56 @@ def test_fixed_random_number(self):
         paddle.enable_static()
 
 
+class TestPoissonFP16OP(TestPoissonOp1):
+    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 TestPoissonBF16Op(OpTest):
+    def setUp(self):
+        self.op_type = "poisson"
+        self.python_api = paddle.tensor.poisson
+        self.__class__.op_type = self.op_type
+        self.config()
+        x = np.full([2048, 1024], self.lam, dtype="float32")
+        out = np.ones([2048, 1024], dtype="float32")
+        self.attrs = {}
+        self.inputs = {'X': convert_float_to_uint16(x)}
+        self.outputs = {'Out': convert_float_to_uint16(out)}
+
+    def config(self):
+        self.lam = 10
+        self.a = 5
+        self.b = 15
+        self.dtype = np.uint16
+
+    def verify_output(self, outs):
+        hist, prob = output_hist(
+            convert_uint16_to_float(np.array(outs[0])), self.lam, self.a, self.b
+        )
+        np.testing.assert_allclose(hist, prob, rtol=0.01)
+
+    def test_check_output(self):
+        place = core.CUDAPlace(0)
+        self.check_output_with_place_customized(self.verify_output, place)
+
+    def test_check_grad(self):
+        place = core.CUDAPlace(0)
+        self.check_grad_with_place(
+            place,
+            ['X'],
+            'Out',
+            user_defined_grads=[np.zeros([2048, 1024], dtype="float32")],
+            user_defined_grad_outputs=[
+                np.random.rand(2048, 1024).astype("float32")
+            ],
+        )
+
+
 if __name__ == "__main__":
     unittest.main()
