Skip to content

Commit

Permalink
add flatten2,reshape2,squueze2_trt_fuse_pass test cast (#41031)
Browse files Browse the repository at this point in the history 
* add flatten2,reshape2,squueze2_trt_fuse_pass  test cast

* add flatten2,reshape2,squueze2_trt_fuse_pass  test cast

* add flatten2,reshape2,squueze2_trt_fuse_pass  test cast
  • Loading branch information
@heliqi
heliqi authored Mar 31, 2022
1 parent e7928a0 commit 7ef6920
Show file tree
Hide file tree
Showing 9 changed files with 477 additions and 133 deletions.
7 changes: 4 additions & 3 deletions paddle/fluid/framework/ir/trt_map_matmul_to_mul_pass.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -498,7 +498,8 @@ void TrtSqueeze2MatmulFusePass::ApplyImpl(ir::Graph* graph) const {
BOOST_GET_CONST(std::vector<int>, squeeze2_op->Op()->GetAttr("axes"));
flag = flag && squeeze2_in_x_rank == 4 &&
squeeze2_op_axes == std::vector<int>{2, 3} &&
(matmul_in_x->outputs).size() == 1;
(matmul_in_x->outputs).size() == 1 &&
matmul_in_y->Var()->Persistable();

bool transpose_X =
BOOST_GET_CONST(bool, matmul_op->Op()->GetAttr("transpose_X"));
Expand Down Expand Up @@ -654,7 +655,7 @@ void TrtReshape2MatmulFusePass::ApplyImpl(ir::Graph* graph) const {
size_t matmul_in_y_rank = (matmul_in_y->Var()->GetShape()).size();
flag = flag && !transpose_X && !transpose_Y &&
std::abs(alpha - 1.0) < 1e-5 && matmul_in_x_rank == 2 &&
matmul_in_y_rank == 2;
matmul_in_y_rank == 2 && matmul_in_y->Var()->Persistable();

std::vector<Node*>& next_ops = matmul_out->outputs;
flag = flag && next_ops.size() == 1 &&
Expand Down Expand Up @@ -740,7 +741,7 @@ void TrtFlatten2MatmulFusePass::ApplyImpl(ir::Graph* graph) const {
size_t matmul_in_y_rank = (matmul_in_y->Var()->GetShape()).size();
pattern_found = pattern_found && !transpose_X && !transpose_Y &&
std::abs(alpha - 1.0) < 1e-5 && matmul_in_x_rank == 2 &&
matmul_in_y_rank == 2;
matmul_in_y_rank == 2 && matmul_in_y->Var()->Persistable();

std::vector<Node*>& next_ops = matmul_out->outputs;
// we further require the matmul op is followed by one elementwise
Expand Down
1 change: 1 addition & 0 deletions paddle/fluid/inference/analysis/ir_pass_manager.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -43,6 +43,7 @@ IRPassManager::IRPassManager(Argument *argument) {
"The scope ptr should not be nullptr."));
graph_->SetNotOwned(framework::ir::kParamScopeAttr, scope_ptr);
}
disable_logs_ = argument->disable_logs();

ARGUMENT_CHECK_FIELD(argument, ir_analysis_passes);
CreatePasses(argument, argument->ir_analysis_passes());
Expand Down
3 changes: 3 additions & 0 deletions python/paddle/fluid/tests/unittests/ir/inference/CMakeLists.txt
View file
Original file line number Diff line number Diff line change
Expand Up @@ -103,6 +103,9 @@ if (WITH_MKLDNN AND TENSORRT_FOUND AND WITH_GPU)
set_tests_properties(test_flatten2_matmul_fuse_pass PROPERTIES TIMEOUT 240)
set_tests_properties(test_squeeze2_matmul_fuse_pass PROPERTIES TIMEOUT 240)
set_tests_properties(test_reshape2_matmul_fuse_pass PROPERTIES TIMEOUT 240)
set_tests_properties(test_trt_flatten2_matmul_fuse_pass PROPERTIES TIMEOUT 240)
set_tests_properties(test_trt_squeeze2_matmul_fuse_pass PROPERTIES TIMEOUT 240)
set_tests_properties(test_trt_reshape2_matmul_fuse_pass PROPERTIES TIMEOUT 240)
set_tests_properties(test_shuffle_channel_detect_pass PROPERTIES TIMEOUT 120)
if (WIN32)
set_tests_properties(test_matmul_scale_fuse_pass PROPERTIES TIMEOUT 300)
Expand Down
38 changes: 0 additions & 38 deletions python/paddle/fluid/tests/unittests/ir/inference/test_flatten2_matmul_fuse_pass.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -39,17 +39,6 @@ class TestFlatten2MatmulFusePass(PassAutoScanTest):
"""

def sample_predictor_configs(self, program_config):
# TRT
# config = self.create_trt_inference_config()
# config.enable_tensorrt_engine(
# max_batch_size=10,
# workspace_size=102400,
# min_subgraph_size=0,
# precision_mode=paddle_infer.PrecisionType.Float32,
# use_static=False,
# use_calib_mode=False)
# yield config, ['mul', 'elementwise_add'], (1e-5, 1e-5)

# cpu
config = self.create_inference_config(use_gpu=False)
yield config, ["mul", "elementwise_add"], (1e-5, 1e-5)
Expand All @@ -58,33 +47,6 @@ def sample_predictor_configs(self, program_config):
config = self.create_inference_config(use_gpu=True)
yield config, ["mul", "elementwise_add"], (1e-5, 1e-5)

def add_ignore_pass_case(self):
# Here we put some skip rules to avoid known bugs
def teller1(program_config, predictor_config):
if predictor_config.tensorrt_engine_enabled():
# On 3080, the results of MatMul and Mul are different
# When the input Y is weight
return True

# On TRT when the input Y is weight, Mul is converted to FC
if "matmul_y" not in program_config.weights \
or "bias" not in program_config.weights:
return True

y_shape = list(program_config.weights["matmul_y"].shape)
bias_shape = program_config.weights["bias"].shape
axis = program_config.ops[2].attrs["axis"]
# bias should be [mul_y_shape[-1]]
if axis == 0 or bias_shape[0] != y_shape[1] or len(
bias_shape) != 1:
return True
return False

self.add_ignore_check_case(
teller1,
IgnoreReasons.PASS_ACCURACY_ERROR,
"The pass error on TRT while shape of bias is not [out_size].", )

def sample_program_config(self, draw):
# 1. Generate shape and attr of flatten2
x_shape = draw(
Expand Down
59 changes: 14 additions & 45 deletions python/paddle/fluid/tests/unittests/ir/inference/test_reshape2_matmul_fuse_pass.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -39,17 +39,6 @@ class TestReshape2MatmulFusePass(PassAutoScanTest):
"""

def sample_predictor_configs(self, program_config):
# TRT
# config = self.create_trt_inference_config()
# config.enable_tensorrt_engine(
# max_batch_size=10,
# workspace_size=102400,
# min_subgraph_size=0,
# precision_mode=paddle_infer.PrecisionType.Float32,
# use_static=False,
# use_calib_mode=False)
# yield config, ['mul', 'elementwise_add'], (1e-5, 1e-5)

# cpu
config = self.create_inference_config(use_gpu=False)
yield config, ["mul", "elementwise_add"], (1e-5, 1e-5)
Expand All @@ -58,33 +47,6 @@ def sample_predictor_configs(self, program_config):
config = self.create_inference_config(use_gpu=True)
yield config, ["mul", "elementwise_add"], (1e-5, 1e-5)

def add_ignore_pass_case(self):
# Here we put some skip rules to avoid known bugs
def teller1(program_config, predictor_config):
if predictor_config.tensorrt_engine_enabled():
# On 3080, the results of MatMul and Mul are different
# When the input Y is weight
return True

# On TRT when the input Y is weight, Mul is converted to FC
if "matmul_y" not in program_config.weights \
or "bias" not in program_config.weights:
return True

y_shape = list(program_config.weights["matmul_y"].shape)
bias_shape = program_config.weights["bias"].shape
axis = program_config.ops[2].attrs["axis"]
# bias should be [mul_y_shape[-1]]
if axis == 0 or bias_shape[0] != y_shape[1] or len(
bias_shape) != 1:
return True
return False

self.add_ignore_check_case(
teller1,
IgnoreReasons.PASS_ACCURACY_ERROR,
"The pass error on TRT while shape of bias is not [out_size].", )

def sample_program_config(self, draw):
# 1. Generate shape and attr of reshape2
reshape = draw(
Expand All @@ -107,14 +69,21 @@ def sample_program_config(self, draw):

# 4. Generate legal attr:axis of elementwise_add
axis = draw(st.integers(min_value=-1, max_value=1))
if axis == 0:
bias_shape = [x_shape[0]]
elif axis == 1:
bias_shape = [y_shape[1]]
else:
bias_shape = [x_shape[0], y_shape[1]]
if axis == 0 or axis == -1:
if draw(st.booleans()):
bias_shape[1] = 1
if axis == 0:
bias_shape = [x_shape[0], ]
else:
bias_shape = [y_shape[1], ]
else:
bias_shape = [x_shape[0], y_shape[1]]
elif axis == 1:
bias_shape = [y_shape[1], ]

if draw(st.integers(min_value=1, max_value=10)) <= 1:
bias_shape[-1] = 1
if len(bias_shape) == 2 and draw(st.booleans()):
bias_shape[0] = 1

reshape2_op = OpConfig(
"reshape2",
Expand Down
57 changes: 10 additions & 47 deletions python/paddle/fluid/tests/unittests/ir/inference/test_squeeze2_matmul_fuse_pass.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -47,45 +47,6 @@ def sample_predictor_configs(self, program_config):
config = self.create_inference_config(use_gpu=True)
yield config, ["mul", "elementwise_add"], (1e-5, 1e-5)

# TRT
# config = self.create_trt_inference_config()
# config.enable_tensorrt_engine(
# max_batch_size=10,
# workspace_size=10240,
# min_subgraph_size=0,
# precision_mode=paddle_infer.PrecisionType.Float32,
# use_static=False,
# use_calib_mode=False)
# yield config, ['mul', 'elementwise_add'], (1e-5, 1e-5)

def add_ignore_pass_case(self):
# Here we put some skip rules to avoid known bugs
def teller1(program_config, predictor_config):
if predictor_config.tensorrt_engine_enabled():
# On 3080, the results of MatMul and Mul are different
# When the input Y is weight
return True

# On TRT when the input Y is weight, Mul is converted to FC
predictor_config.exp_disable_tensorrt_ops(["elementwise_add"])
if "matmul_y" not in program_config.weights \
or "bias" not in program_config.weights:
return True

y_shape = list(program_config.weights["matmul_y"].shape)
bias_shape = program_config.weights["bias"].shape
axis = program_config.ops[2].attrs["axis"]
# bias should be [mul_y_shape[-1]]
if axis == 0 or bias_shape[0] != y_shape[1] or len(
bias_shape) != 1:
return True
return False

self.add_ignore_check_case(
teller1,
IgnoreReasons.PASS_ACCURACY_ERROR,
"The pass error on TRT while shape of bias is not [out_size].", )

def sample_program_config(self, draw):
# 1. Generate shape of input:X of squeeze2
x_shape = draw(
Expand All @@ -111,19 +72,21 @@ def sample_program_config(self, draw):
# 4. Generate legal attr:axis of elementwise_add
axis = draw(st.integers(min_value=-1, max_value=1))
if axis == 0 or axis == -1:
bias_shape = [x_shape[0], y_shape[1]]
else:
if draw(st.booleans()):
if axis == 0:
bias_shape = [x_shape[0], ]
else:
bias_shape = [y_shape[1], ]
else:
bias_shape = [x_shape[0], y_shape[1]]
elif axis == 1:
bias_shape = [y_shape[1], ]
if draw(st.booleans()):

if draw(st.integers(min_value=1, max_value=10)) <= 1:
bias_shape[-1] = 1
if len(bias_shape) == 2 and draw(st.booleans()):
bias_shape[0] = 1

axis = 0
bias_shape = [2, ]
x_shape = [2, 1, 1, 1]
y_shape = [1, 2]

squeeze2_op = OpConfig(
"squeeze2",
inputs={"X": ["squeeze2_x"], },
Expand Down
Loading

0 comments on commit 7ef6920

Please sign in to comment.