Skip to content

Commit 840a079

Browse files
b3602sssb3602sss
committed
more uts
1 parent 7033b3a commit 840a079

File tree

3 files changed

+138
-2
lines changed

3 files changed

+138
-2
lines changed

‎paddle/fluid/inference/tensorrt/convert/reduce_op.cc‎

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -72,6 +72,7 @@ class ReduceOpConverter : public OpConverter {
7272
if (x < 0) {
7373
res |= 1 << (x + input_dims);
7474
} else {
75+
if (!engine_->with_dynamic_shape()) x = x - 1;
7576
res |= 1 << x;
7677
}
7778
}

‎paddle/fluid/inference/tensorrt/op_teller.cc‎

Lines changed: 10 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -718,6 +718,16 @@ bool OpTeller::Tell(const framework::ir::Node* node, bool use_no_calib_int8,
718718
"reduce_all)";
719719
return false;
720720
}
721+
722+
// The batch size dimension cannot be reduced if it's not dynamic shape.
723+
if (!with_dynamic_shape) {
724+
if (desc.HasAttr("reduce_all")) return false;
725+
std::vector<int32_t> dim =
726+
BOOST_GET_CONST(std::vector<int32_t>, desc.GetAttr("dim"));
727+
for (auto x : dim) {
728+
if (!x) return false;
729+
}
730+
}
721731
}
722732

723733
if ((*teller)(op_type, desc, use_no_calib_int8)) return true;

‎python/paddle/fluid/tests/unittests/ir/inference/test_trt_reduce_mean_op.py‎

Lines changed: 127 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -27,7 +27,7 @@ class TRTReduceMeanTest(InferencePassTest):
2727
def setUp(self):
2828
with fluid.program_guard(self.main_program, self.startup_program):
2929
data = fluid.data(
30-
name="data", shape=[-1, 3, 224, 224], dtype="float32")
30+
name="data", shape=[-1, 3, -1, -1], dtype="float32")
3131
reduce_mean = fluid.layers.reduce_mean(
3232
data, dim=[2, -1], keep_dim=True)
3333
out = fluid.layers.batch_norm(reduce_mean, is_test=True)
@@ -40,7 +40,35 @@ def setUp(self):
4040
1 << 30, 32, 1, AnalysisConfig.Precision.Float32, False, False)
4141
self.fetch_list = [out]
4242
self.dynamic_shape_params = TRTReduceMeanTest.DynamicShapeParam({
43-
'data': [1, 3, 224, 224]
43+
'data': [1, 3, 64, 64]
44+
}, {'data': [3, 3, 224, 224]}, {'data': [3, 3, 224, 224]}, False)
45+
46+
def test_check_output(self):
47+
if core.is_compiled_with_cuda():
48+
use_gpu = True
49+
self.check_output_with_option(use_gpu, flatten=True)
50+
self.assertTrue(
51+
PassVersionChecker.IsCompatible('tensorrt_subgraph_pass'))
52+
53+
54+
class TRTReduceMeanTestFP16(InferencePassTest):
55+
def setUp(self):
56+
with fluid.program_guard(self.main_program, self.startup_program):
57+
data = fluid.data(
58+
name="data", shape=[-1, 3, -1, -1], dtype="float32")
59+
reduce_mean = fluid.layers.reduce_mean(
60+
data, dim=[2, -1], keep_dim=True)
61+
out = fluid.layers.batch_norm(reduce_mean, is_test=True)
62+
63+
self.feeds = {
64+
"data": np.random.random([3, 3, 224, 224]).astype("float32"),
65+
}
66+
self.enable_trt = True
67+
self.trt_parameters = TRTReduceMeanTestFP16.TensorRTParam(
68+
1 << 30, 32, 1, AnalysisConfig.Precision.Half, False, False)
69+
self.fetch_list = [out]
70+
self.dynamic_shape_params = TRTReduceMeanTestFP16.DynamicShapeParam({
71+
'data': [1, 3, 64, 64]
4472
}, {'data': [3, 3, 224, 224]}, {'data': [3, 3, 224, 224]}, False)
4573

4674
def test_check_output(self):
@@ -78,5 +106,102 @@ def test_check_output(self):
78106
PassVersionChecker.IsCompatible('tensorrt_subgraph_pass'))
79107

80108

109+
class TRTReduceMeanTestStatic(InferencePassTest):
110+
def setUp(self):
111+
with fluid.program_guard(self.main_program, self.startup_program):
112+
data = fluid.data(
113+
name="data", shape=[3, 3, 224, 224], dtype="float32")
114+
reduce_mean = fluid.layers.reduce_mean(
115+
data, dim=[2, -1], keep_dim=True)
116+
out = fluid.layers.batch_norm(reduce_mean, is_test=True)
117+
118+
self.feeds = {
119+
"data": np.random.random([3, 3, 224, 224]).astype("float32"),
120+
}
121+
self.enable_trt = True
122+
self.trt_parameters = TRTReduceMeanTestStatic.TensorRTParam(
123+
1 << 30, 32, 1, AnalysisConfig.Precision.Float32, False, False)
124+
self.fetch_list = [out]
125+
126+
def test_check_output(self):
127+
if core.is_compiled_with_cuda():
128+
use_gpu = True
129+
self.check_output_with_option(use_gpu, flatten=True)
130+
self.assertTrue(
131+
PassVersionChecker.IsCompatible('tensorrt_subgraph_pass'))
132+
133+
134+
class TRTReduceMeanStaticAllTest(InferencePassTest):
135+
def setUp(self):
136+
with fluid.program_guard(self.main_program, self.startup_program):
137+
data = fluid.data(
138+
name="data", shape=[4, 3, 224, 224], dtype="float32")
139+
reduce_mean = fluid.layers.reduce_mean(data, keep_dim=True)
140+
out = fluid.layers.batch_norm(reduce_mean, is_test=True)
141+
142+
self.feeds = {
143+
"data": np.random.random([4, 3, 224, 224]).astype("float32"),
144+
}
145+
self.enable_trt = True
146+
self.trt_parameters = TRTReduceMeanStaticAllTest.TensorRTParam(
147+
1 << 30, 32, 1, AnalysisConfig.Precision.Float32, False, False)
148+
self.fetch_list = [out]
149+
150+
def test_check_output(self):
151+
if core.is_compiled_with_cuda():
152+
use_gpu = True
153+
self.check_output_with_option(use_gpu, flatten=True)
154+
self.assertTrue(
155+
PassVersionChecker.IsCompatible('tensorrt_subgraph_pass'))
156+
157+
158+
class TRTReduceMeanStaticFP16(InferencePassTest):
159+
def setUp(self):
160+
with fluid.program_guard(self.main_program, self.startup_program):
161+
data = fluid.data(
162+
name="data", shape=[4, 3, 224, 224], dtype="float32")
163+
reduce_mean = fluid.layers.reduce_mean(data, keep_dim=True)
164+
out = fluid.layers.batch_norm(reduce_mean, is_test=True)
165+
166+
self.feeds = {
167+
"data": np.random.random([4, 3, 224, 224]).astype("float32"),
168+
}
169+
self.enable_trt = True
170+
self.trt_parameters = TRTReduceMeanStaticFP16.TensorRTParam(
171+
1 << 30, 32, 1, AnalysisConfig.Precision.Half, False, False)
172+
self.fetch_list = [out]
173+
174+
def test_check_output(self):
175+
if core.is_compiled_with_cuda():
176+
use_gpu = True
177+
self.check_output_with_option(use_gpu, flatten=True)
178+
self.assertTrue(
179+
PassVersionChecker.IsCompatible('tensorrt_subgraph_pass'))
180+
181+
182+
class TRTReduceMeanFP16Static(InferencePassTest):
183+
def setUp(self):
184+
with fluid.program_guard(self.main_program, self.startup_program):
185+
data = fluid.data(
186+
name="data", shape=[4, 3, 224, 224], dtype="float32")
187+
reduce_mean = fluid.layers.reduce_mean(data, keep_dim=True)
188+
out = fluid.layers.batch_norm(reduce_mean, is_test=True)
189+
190+
self.feeds = {
191+
"data": np.random.random([4, 3, 224, 224]).astype("float32"),
192+
}
193+
self.enable_trt = True
194+
self.trt_parameters = TRTReduceMeanFP16Static.TensorRTParam(
195+
1 << 30, 32, 1, AnalysisConfig.Precision.Half, True, False)
196+
self.fetch_list = [out]
197+
198+
def test_check_output(self):
199+
if core.is_compiled_with_cuda():
200+
use_gpu = True
201+
self.check_output_with_option(use_gpu, flatten=True)
202+
self.assertTrue(
203+
PassVersionChecker.IsCompatible('tensorrt_subgraph_pass'))
204+
205+
81206
if __name__ == "__main__":
82207
unittest.main()

0 commit comments

Comments
 (0)