[Prim] fix clip_decomp with where instead of maximum

paddle/fluid/primitive/decomp_rule/decomp_rule/composite.h

@@ -931,17 +931,17 @@ Tensor clip_decomp(const Tensor& x, const Tensor& min, const Tensor& max) {
  auto min_reshape = min;
  auto max_reshape = max;
 
- if (x.shape().size() == 0) {
- min_reshape = reshape<T>(min_reshape, {});
- max_reshape = reshape<T>(max_reshape, {});
- }
-
  if (has_dynamic_shape(x.shape())) {
  min_reshape = backend::expand<T>(min_reshape, shape64<T>(x));
  max_reshape = backend::expand<T>(max_reshape, shape64<T>(x));
  } else {
- min_reshape = expand<T>(min_reshape, x.shape());
- max_reshape = expand<T>(max_reshape, x.shape());
+ if (x.shape().size() == 0) {
+ min_reshape = reshape<T>(min_reshape, {});
+ max_reshape = reshape<T>(max_reshape, {});
+ } else {
+ min_reshape = expand<T>(min_reshape, x.shape());
+ max_reshape = expand<T>(max_reshape, x.shape());
+ }
  }
  if (min_reshape.dtype() != x.dtype()) {
  min_reshape = cast<T>(min_reshape, x.dtype());
@@ -950,8 +950,9 @@ Tensor clip_decomp(const Tensor& x, const Tensor& min, const Tensor& max) {
  if (max_reshape.dtype() != x.dtype()) {
  max_reshape = cast<T>(max_reshape, x.dtype());
  }
-
- auto ans = maximum<T>(minimum<T>(x, max_reshape), min_reshape);
+ auto ans = where<T>(x <= max_reshape,
+ where<T>(x >= min_reshape, x, min_reshape),
+ max_reshape);
  return ans;
 }
 

test/legacy_test/test_clip_op.py

@@ -46,7 +46,7 @@ def setUp(self):
  else:
  max_v = self.attrs['max']
 
- input = np.random.random(self.shape).astype(self.dtype)
+ input = self.generate_input()
  input[np.abs(input - min_v) < self.max_relative_error] = 0.5
  input[np.abs(input - max_v) < self.max_relative_error] = 0.5
  self.inputs['X'] = input
@@ -67,7 +67,7 @@ def test_check_output(self):
 
  def test_check_grad_normal(self):
  paddle.enable_static()
- self.check_grad(['X'], 'Out', check_pir=True)
+ self.check_grad(['X'], 'Out', check_pir=True, check_prim_pir=True)
  paddle.disable_static()
 
  def initTestCase(self):
@@ -78,6 +78,9 @@ def initTestCase(self):
  self.inputs['Max'] = np.array([0.8]).astype(self.dtype)
  self.inputs['Min'] = np.array([0.1]).astype(self.dtype)
 
+ def generate_input(self):
+ return np.random.random(self.shape).astype(self.dtype)
+
 
 class TestCase1(TestClipOp):
  def initTestCase(self):
@@ -121,6 +124,19 @@ def initTestCase(self):
  self.min = 0.5
 
 
+class TestCase6(TestClipOp):
+ def initTestCase(self):
+ self.dtype = np.float32
+ self.shape = (4, 8, 16)
+ self.max = 1.0
+ self.min = 0.5
+
+ def generate_input(self):
+ return np.random.choice([self.min, self.max], self.shape).astype(
+ self.dtype
+ )
+
+
 class TestFP16Case1(TestClipOp):
  def initTestCase(self):
  self.dtype = np.float16
@@ -163,6 +179,19 @@ def initTestCase(self):
  self.min = 0.5
 
 
+class TestFP16Case6(TestClipOp):
+ def initTestCase(self):
+ self.dtype = np.float16
+ self.shape = (4, 8, 16)
+ self.max = 1.0
+ self.min = 0.5
+
+ def generate_input(self):
+ return np.random.choice([self.min, self.max], self.shape).astype(
+ self.dtype
+ )
+
+
 @unittest.skipIf(
  not core.is_compiled_with_cuda()
  or not core.is_bfloat16_supported(core.CUDAPlace(0)),