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

[inference][trt] upgrade prelu op #48528

Merged
merged 3 commits into from
Dec 9, 2022
Merged

[inference][trt] upgrade prelu op #48528

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
99bf8ba
add prelu
zhangjun Nov 29, 2022
b415451
add prelu
zhangjun Nov 29, 2022
6d04c4d
fix channel mode contains the batch
zhangjun Dec 2, 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
113 changes: 75 additions & 38 deletions paddle/fluid/inference/tensorrt/convert/prelu_op.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -31,59 +31,96 @@ class PReluOpConverter : public OpConverter {

framework::OpDesc op_desc(op, nullptr);
// Declare inputs
size_t input_num = op_desc.Input("X").size();
auto* input = engine_->GetITensor(op_desc.Input("X")[0]);
auto input_dims = input->getDimensions();
// Get attrs
std::string mode = PADDLE_GET_CONST(std::string, op_desc.GetAttr("mode"));
std::string data_format = "NCHW";
if (op_desc.HasAttr("data_format")) {
data_format =
PADDLE_GET_CONST(std::string, op_desc.GetAttr("data_format"));
}
auto* alpha_var = scope.FindVar(op_desc.Input("Alpha")[0]);
auto* alpha_tensor = alpha_var->GetMutable<phi::DenseTensor>();

auto alpha_weight =
engine_->GetFp32TrtWeight(op_desc.Input("Alpha")[0], *alpha_tensor);
auto* alpha_var = scope.FindVar(op_desc.Input("Alpha")[0]);
auto* alpha_weight = alpha_var->GetMutable<phi::DenseTensor>();
auto w_dims = alpha_weight->dims();
auto alpha_data =
engine_->GetFp32TrtWeight(op_desc.Input("Alpha")[0], *alpha_weight);

platform::CPUPlace cpu_place;
nvinfer1::Dims trt_w_dims;
trt_w_dims.nbDims = w_dims.size();
for (int i = 0; i < trt_w_dims.nbDims; i++) {
trt_w_dims.d[i] = w_dims[i];
}

nvinfer1::ILayer* layer = nullptr;
if (engine_->with_dynamic_shape()) {
plugin::PReluPluginDynamic* plugin = new plugin::PReluPluginDynamic(
static_cast<const float*>(alpha_weight.get().values),
alpha_tensor->numel(),
mode,
data_format);
layer = engine_->AddDynamicPlugin(&input, input_num, plugin);
} else {
#if IS_TRT_VERSION_GE(7000)
nvinfer1::Dims dims;
dims.nbDims = 0;
// jump batch dim
for (int i = 1; i < alpha_tensor->dims().size(); i++) {
dims.d[dims.nbDims++] = alpha_tensor->dims()[i];
}
for (; dims.nbDims < input->getDimensions().nbDims; dims.nbDims++) {
dims.d[dims.nbDims] = 1;
// The `element` or `channel` mode contains the batch using static shape.
if ((mode == "element" || mode == "channel") &&
!engine_->with_dynamic_shape() &&
(trt_w_dims.nbDims - 1 == input_dims.nbDims)) {
trt_w_dims.nbDims--;
for (int i = 0; i < trt_w_dims.nbDims; i++) {
trt_w_dims.d[i] = trt_w_dims.d[i + 1];
}
}

auto alpha_layer =
TRT_ENGINE_ADD_LAYER(engine_, Constant, dims, alpha_weight.get());
auto alpha_layer_output = alpha_layer->getOutput(0);

layer = TRT_ENGINE_ADD_LAYER(
engine_, ParametricReLU, *input, *alpha_layer_output);
#else
plugin::PReluPlugin* plugin = new plugin::PReluPlugin(
static_cast<const float*>(alpha_weight.get().values),
alpha_tensor->numel(),
mode,
data_format);
layer = engine_->AddPlugin(&input, input_num, plugin);
#endif
nvinfer1::ITensor* alpha_tensor =
TRT_ENGINE_ADD_LAYER(engine_, Constant, trt_w_dims, alpha_data.get())
->getOutput(0);

auto alpha_dims = alpha_tensor->getDimensions();
nvinfer1::ITensor* real_alpha_tensor = alpha_tensor;
if (alpha_dims.nbDims != input_dims.nbDims) {
auto* reshape_layer =
TRT_ENGINE_ADD_LAYER(engine_, Shuffle, *alpha_tensor);
int c = alpha_dims.d[0];
if (engine_->with_dynamic_shape()) {
std::vector<nvinfer1::ITensor*> itensors;
auto* n_tensor = Add1DConstantLayer(1);
auto* c_tensor = Add1DConstantLayer(c);
nvinfer1::ITensor* hw_tensor = nullptr;
nvinfer1::ITensor* shape_tensor = nullptr;
if (input_dims.nbDims - 2 > 0) {
hw_tensor = Add1DConstantLayer(
std::vector<int32_t>(input_dims.nbDims - 2, 1));
}
if (data_format == "NCHW") {
if (hw_tensor != nullptr) {
shape_tensor = Concat(
std::vector<nvinfer1::ITensor*>{n_tensor, c_tensor, hw_tensor});

} else {
shape_tensor =
Concat(std::vector<nvinfer1::ITensor*>{n_tensor, c_tensor});
}
} else {
if (hw_tensor != nullptr) {
shape_tensor = Concat(
std::vector<nvinfer1::ITensor*>{n_tensor, hw_tensor, c_tensor});
} else {
shape_tensor =
Concat(std::vector<nvinfer1::ITensor*>{n_tensor, c_tensor});
}
}
reshape_layer->setInput(1, *shape_tensor);
} else {
nvinfer1::Dims reshape_dim;
reshape_dim.nbDims = input_dims.nbDims;
std::fill(reshape_dim.d, reshape_dim.d + input_dims.nbDims, 1);
if (data_format == "NCHW") {
reshape_dim.d[0] = c;
} else if (data_format == "NHWC") {
reshape_dim.d[input_dims.nbDims - 1] = c;
}
reshape_layer->setReshapeDimensions(reshape_dim);
}
real_alpha_tensor = reshape_layer->getOutput(0);
}

nvinfer1::ILayer* layer = nullptr;

layer = TRT_ENGINE_ADD_LAYER(
engine_, ParametricReLU, *input, *real_alpha_tensor);

auto output_name = op_desc.Output("Out")[0];
RreplenishLayerAndOutput(layer, "prelu", {output_name}, test_mode);
}
Expand Down
12 changes: 6 additions & 6 deletions python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_prelu.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -49,22 +49,22 @@ def generate_alpha(attrs: List[Dict[str, Any]], dim1, dim2, dim3):
if dim1 != 0:
shape.append(dim1)
if dim2 != 0:
shape.append(1)
shape.append(dim2)
if dim3 != 0:
shape.append(1)
return np.random.random(size=shape).astype(np.float32)
shape.append(dim3)
return np.random.random(size=shape[1]).astype(np.float32)
elif (
attrs[0]["mode"] == "channel"
and attrs[0]["data_format"] == "NHWC"
):
shape = [1]
if dim1 != 0:
shape.append(1)
shape.append(dim1)
if dim2 != 0:
shape.append(1)
shape.append(dim2)
if dim3 != 0:
shape.append(dim3)
return np.random.random(size=shape).astype(np.float32)
return np.random.random(size=shape[-1]).astype(np.float32)
elif attrs[0]["mode"] == "element":
shape = [1]
if dim1 != 0:
Expand Down