Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

[inference] update trt convert reduce op&ut,test=develop #39088

Merged
merged 12 commits into from
Jan 25, 2022
Merged

[inference] update trt convert reduce op&ut,test=develop #39088

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
12 commits
Select commit Hold shift + click to select a range
7052f85
[inference] update convert reduce op&ut,test=develop
zhangjun Jan 20, 2022
9b7db2b
update
zhangjun Jan 20, 2022
5ca9ee9
update
zhangjun Jan 20, 2022
c030eb7
update
zhangjun Jan 20, 2022
fec3c5b
update
zhangjun Jan 20, 2022
132a0eb
add int32 support
zhangjun Jan 21, 2022
82bfef7
add int32 support
zhangjun Jan 21, 2022
eaf47d1
add comments
zhangjun Jan 21, 2022
f0fa5b8
trt < 7.0 do not support int32
zhangjun Jan 24, 2022
2dbbe96
test=develop
zhangjun Jan 24, 2022
538cb01
update
zhangjun Jan 24, 2022
d8b47c7
test=develop
zhangjun Jan 24, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 2 additions & 0 deletions paddle/fluid/inference/tensorrt/convert/reduce_op.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -83,6 +83,8 @@ class ReduceOpConverter : public OpConverter {
}

auto output_name = op_desc.Output("Out")[0];
// Ensure that the output type and input type are consistent.
layer->getOutput(0)->setType(layer->getInput(0)->getType());
RreplenishLayerAndOutput(layer, op_type, {output_name}, test_mode);
}

Expand Down
34 changes: 26 additions & 8 deletions paddle/fluid/inference/tensorrt/op_teller.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1490,30 +1490,48 @@ bool OpTeller::Tell(const framework::ir::Node* node, bool use_no_calib_int8,
VLOG(3) << "the " << op_type
<< " does not have attr (keep_dim or dim or "
"reduce_all)";
std::cout << "attr " << desc.HasAttr("keep_dim") << " "
<< desc.HasAttr("dim") << " " << desc.HasAttr("reduce_all");
return false;
}

auto* block = desc.Block();
if (block == nullptr) {
VLOG(3) << "The block desc is nullptr, we can't continue to analyze. "
"Developers need to check whether block_desc is passed in "
"the pass.";
return false;
}

// The batch size dimension cannot be reduced if it's not dynamic shape.
auto* x_var_desc = block->FindVar(desc.Input("X")[0]);
if (!with_dynamic_shape) {
if (BOOST_GET_CONST(bool, desc.GetAttr("reduce_all"))) return false;
std::vector<int32_t> dim =
BOOST_GET_CONST(std::vector<int32_t>, desc.GetAttr("dim"));
const auto input_shape = x_var_desc->GetShape();
for (auto x : dim) {
if (!x) return false;
if (x == 0 || (x + input_shape.size() == 0)) return false;
}

} else {
if (BOOST_GET_CONST(bool, desc.GetAttr("reduce_all")) &&
!BOOST_GET_CONST(bool, desc.GetAttr("keep_dim")))
return false;
}
if (desc.HasAttr("out_dtype")) {
int out_dtype = BOOST_GET_CONST(int32_t, desc.GetAttr("out_dtype"));
if (out_dtype != -1) {
return false;
}

auto dtype = x_var_desc->GetDataType();
#if IS_TRT_VERSION_GE(7000)
if (dtype != framework::proto::VarType::INT32 &&
dtype != framework::proto::VarType::FP32) {
zhangjun marked this conversation as resolved.
Show resolved Hide resolved
VLOG(3) << "reduce op input data type must be int32 or float32";
return false;
}
#else
if (dtype != framework::proto::VarType::FP32) {
VLOG(3)
<< "reduce op input data type must be float32 using TensorRT < 7.0";
return false;
}
#endif
}
#if IS_TRT_VERSION_GE(7000)
if (op_type == "tile") {
Expand Down
4 changes: 3 additions & 1 deletion paddle/fluid/operators/reduce_ops/reduce_mean_op.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -22,4 +22,6 @@ REGISTER_OP_CUDA_KERNEL(
ops::ReduceCudaKernel<paddle::platform::float16, kps::AddFunctor,
kps::DivideFunctor>,
ops::ReduceCudaKernel<float, kps::AddFunctor, kps::DivideFunctor>,
ops::ReduceCudaKernel<double, kps::AddFunctor, kps::DivideFunctor>);
ops::ReduceCudaKernel<double, kps::AddFunctor, kps::DivideFunctor>,
ops::ReduceCudaKernel<int, kps::AddFunctor, kps::DivideFunctor>,
ops::ReduceCudaKernel<int64_t, kps::AddFunctor, kps::DivideFunctor>);
zhangjun marked this conversation as resolved.
Show resolved Hide resolved
36 changes: 16 additions & 20 deletions python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_reduce_mean.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -36,26 +36,32 @@ def is_program_valid(self, program_config: ProgramConfig) -> bool:
return False
if len(attrs[0]["dim"]) == 0:
return False
## skip not use
if attrs[0]["out_dtype"] != -1:
return False

ver = paddle_infer.get_trt_compile_version()
if ver[0] * 1000 + ver[1] * 100 + ver[0] * 10 < 7000:
if attrs[0]['out_dtype'] == 2:
return False

return True

def sample_program_configs(self):
def generate_input1(attrs: List[Dict[str, Any]]):
return np.random.random([1, 3, 64, 64]).astype(np.float32)
def generate_input1(dtype, attrs: List[Dict[str, Any]]):
if dtype == -1 or dtype == 5:
return np.random.random([1, 3, 64, 64]).astype(np.float32)
elif dtype == 2:
return np.random.random([1, 3, 64, 64]).astype(np.int32)

for keep_dim in [False, True]:
for keep_dim in [True, False]:
for dim in [[], [1], [0], [0, 1], [1, 2, 3], [-2, 0, 3], [-3],
[-4, 1], [3, 4, 5]]:
for reduce_all in [False, True]:
for out_dtype in [-1, 0, 1]:
for reduce_all in [True, False]:
for out_dtype in [-1, 2, 5]:
dics = [{
"keep_dim": keep_dim,
"dim": dim,
"reduce_all": reduce_all,
"out_dtype": out_dtype
"out_dtype": out_dtype,
"in_dtype": out_dtype,
}, {}]

ops_config = [{
Expand All @@ -75,7 +81,7 @@ def generate_input1(attrs: List[Dict[str, Any]]):
weights={},
inputs={
"input_data": TensorConfig(data_gen=partial(
generate_input1, dics))
generate_input1, out_dtype, dics))
},
outputs=["reduce_output_data"])

Expand Down Expand Up @@ -134,16 +140,6 @@ def generate_trt_nodes_num(attrs, dynamic_shape):
pass

def add_skip_trt_case(self):
def teller1(program_config, predictor_config):
if program_config.ops[0].attrs['out_dtype'] != -1:
return True
return False

self.add_skip_case(
teller1, SkipReasons.TRT_NOT_IMPLEMENTED,
"NOT Implemented: we will add out_dtype not equal to -1 in the future"
)

pass

def test(self):
Expand Down
31 changes: 11 additions & 20 deletions python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_reduce_sum.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -37,26 +37,27 @@ def is_program_valid(self, program_config: ProgramConfig) -> bool:
return False
if len(attrs[0]["dim"]) == 0:
return False
## skip not use
if attrs[0]["out_dtype"] != -1:
return False

return True

def sample_program_configs(self):
def generate_input1(attrs: List[Dict[str, Any]]):
return np.random.random([1, 3, 64, 64]).astype(np.float32)
def generate_input1(dtype, attrs: List[Dict[str, Any]]):
if dtype == -1 or dtype == 5:
return np.random.random([1, 3, 64, 64]).astype(np.float32)
elif dtype == 2:
return np.random.random([1, 3, 64, 64]).astype(np.int32)

for keep_dim in [False, True]:
for keep_dim in [True, False]:
for dim in [[], [1], [0], [0, 1], [1, 2, 3], [-2, 0, 3], [-3],
[-4, 1], [3, 4, 5]]:
for reduce_all in [False, True]:
for out_dtype in [-1, 0, 1]:
for reduce_all in [True, False]:
for out_dtype in [-1, 2, 5]:
dics = [{
"keep_dim": keep_dim,
"dim": dim,
"reduce_all": reduce_all,
"out_dtype": out_dtype
"out_dtype": out_dtype,
"in_dtype": out_dtype,
}, {}]

ops_config = [{
Expand All @@ -76,7 +77,7 @@ def generate_input1(attrs: List[Dict[str, Any]]):
weights={},
inputs={
"input_data": TensorConfig(data_gen=partial(
generate_input1, dics))
generate_input1, out_dtype, dics))
},
outputs=["reduce_output_data"])

Expand Down Expand Up @@ -134,16 +135,6 @@ def generate_trt_nodes_num(attrs, dynamic_shape):
pass

def add_skip_trt_case(self):
def teller1(program_config, predictor_config):
if program_config.ops[0].attrs['out_dtype'] != -1:
return True
return False

self.add_skip_case(
teller1, SkipReasons.TRT_NOT_IMPLEMENTED,
"NOT Implemented: we will add out_dtype not equal to -1 in the future"
)

pass

def test(self):
Expand Down