COMPMID-617 Add validation to NEConvolutionLayer
Change-Id: I796a13e6ea672e274aaa8234ee0689828fec7292
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/111348
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Ioan-Cristian Szabo <ioan-cristian.szabo@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
diff --git a/src/core/NEON/kernels/NEGEMMMatrixMultiplyKernel.cpp b/src/core/NEON/kernels/NEGEMMMatrixMultiplyKernel.cpp
index a583c1d..aa5e2dd 100644
--- a/src/core/NEON/kernels/NEGEMMMatrixMultiplyKernel.cpp
+++ b/src/core/NEON/kernels/NEGEMMMatrixMultiplyKernel.cpp
@@ -1408,6 +1408,129 @@
},
ina, inb, out);
}
+
+Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output)
+{
+ ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::F16, DataType::F32, DataType::QS8, DataType::QS16);
+ ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1, output);
+ ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input0, input1, output);
+ ARM_COMPUTE_UNUSED(input0);
+ ARM_COMPUTE_UNUSED(input1);
+ ARM_COMPUTE_UNUSED(output);
+
+ if(output->dimension(1) == 1)
+ {
+ ARM_COMPUTE_RETURN_ERROR_ON(input0->dimension(0) != input1->dimension(1));
+ }
+
+ return Status{};
+}
+
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input0, ITensorInfo *input1, ITensorInfo *output)
+{
+ Window win = Window();
+ bool window_changed = false;
+
+ unsigned int num_elems_processed_per_iteration_x = 0;
+ const unsigned int num_elems_processed_per_iteration_y = 4;
+
+ // Check if the output tensor is a vector. If so,the kernel runs the vector-matrix multiplication
+ if((output->dimension(1) == 1))
+ {
+ switch(input0->data_type())
+ {
+ case DataType::F32:
+ {
+ num_elems_processed_per_iteration_x = 16;
+ break;
+ }
+ case DataType::QS8:
+ {
+ num_elems_processed_per_iteration_x = 32;
+ break;
+ }
+ case DataType::QS16:
+ {
+ num_elems_processed_per_iteration_x = 16;
+ break;
+ }
+#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+ case DataType::F16:
+ {
+ num_elems_processed_per_iteration_x = 32;
+ break;
+ }
+#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
+ default:
+ {
+ ARM_COMPUTE_ERROR("Data type not supported");
+ break;
+ }
+ }
+
+ // Configure kernel window
+ win = calculate_max_window(*output, Steps(num_elems_processed_per_iteration_x));
+
+ AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration_x);
+
+ window_changed = update_window_and_padding(win,
+ AccessWindowStatic(input0, 0, 0, input0->tensor_shape().x(), 1),
+ AccessWindowHorizontal(input1, 0, num_elems_processed_per_iteration_x),
+ output_access);
+
+ Coordinates coord;
+ coord.set_num_dimensions(output->num_dimensions());
+ output_access.set_valid_region(win, ValidRegion(coord, output->tensor_shape()));
+ }
+ else
+ {
+ switch(input0->data_type())
+ {
+ case DataType::F32:
+ {
+ num_elems_processed_per_iteration_x = 8;
+ break;
+ }
+ case DataType::QS8:
+ {
+ num_elems_processed_per_iteration_x = 32;
+ break;
+ }
+ case DataType::QS16:
+ {
+ num_elems_processed_per_iteration_x = 8;
+ break;
+ }
+#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+ case DataType::F16:
+ {
+ num_elems_processed_per_iteration_x = 8;
+ break;
+ }
+#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
+ default:
+ {
+ ARM_COMPUTE_ERROR("Data type not supported");
+ break;
+ }
+ }
+
+ // Configure kernel window
+ win = calculate_max_window(*output, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
+
+ AccessWindowRectangle output_access(output, 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y);
+
+ window_changed = update_window_and_padding(win,
+ AccessWindowRectangle(input0, 0, 0, 4, 1, 1.f, 0.25f),
+ AccessWindowStatic(input1, 0, 0, input1->tensor_shape().x(), ceil_to_multiple(input1->tensor_shape().y(), 4)),
+ output_access);
+
+ output_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), output->tensor_shape()));
+ }
+
+ Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
+ return std::make_pair(err, win);
+}
} // namespace
NEGEMMMatrixMultiplyKernel::NEGEMMMatrixMultiplyKernel()
@@ -1417,120 +1540,27 @@
void NEGEMMMatrixMultiplyKernel::configure(const ITensor *input0, const ITensor *input1, ITensor *output, float alpha)
{
- ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::F16, DataType::F32, DataType::QS8, DataType::QS16);
- ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1, output);
- ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input0, input1, output);
-
- if(output->info()->dimension(1) == 1)
- {
- ARM_COMPUTE_ERROR_ON(input0->info()->dimension(0) != input1->info()->dimension(1));
- }
+ // Perform validate step
+ ARM_COMPUTE_ERROR_ON_NULLPTR(input0, input1, output);
+ ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input0->info(), input1->info(), output->info()));
_input0 = input0;
_input1 = input1;
_output = output;
_alpha = alpha;
- unsigned int num_elems_processed_per_iteration_x = 0;
- const unsigned int num_elems_processed_per_iteration_y = 4;
+ // Configure kernel window
+ auto win_config = validate_and_configure_window(input0->info(), input1->info(), output->info());
+ ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
+ INEKernel::configure(win_config.second);
+}
- // Check if the output tensor is a vector. If so,the kernel runs the vector-matrix multiplication
- if((output->info()->dimension(1) == 1))
- {
- switch(input0->info()->data_type())
- {
- case DataType::F32:
- {
- num_elems_processed_per_iteration_x = 16;
- break;
- }
- case DataType::QS8:
- {
- num_elems_processed_per_iteration_x = 32;
- break;
- }
- case DataType::QS16:
- {
- num_elems_processed_per_iteration_x = 16;
- break;
- }
-#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
- case DataType::F16:
- {
- num_elems_processed_per_iteration_x = 32;
- break;
- }
-#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
- default:
- {
- ARM_COMPUTE_ERROR("Data type not supported");
- break;
- }
- }
+Status NEGEMMMatrixMultiplyKernel::validate(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output)
+{
+ ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input0, input1, output));
+ ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input0->clone().get(), input1->clone().get(), output->clone().get()).first);
- // Configure kernel window
- Window win = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration_x));
-
- AccessWindowHorizontal output_access(output->info(), 0, num_elems_processed_per_iteration_x);
-
- update_window_and_padding(win,
- AccessWindowStatic(input0->info(), 0, 0, input0->info()->tensor_shape().x(), 1),
- AccessWindowHorizontal(input1->info(), 0, num_elems_processed_per_iteration_x),
- output_access);
-
- Coordinates coord;
- coord.set_num_dimensions(output->info()->num_dimensions());
- output_access.set_valid_region(win, ValidRegion(coord, output->info()->tensor_shape()));
-
- INEKernel::configure(win);
- }
- else
- {
- switch(input0->info()->data_type())
- {
- case DataType::F32:
- {
- num_elems_processed_per_iteration_x = 8;
- break;
- }
- case DataType::QS8:
- {
- num_elems_processed_per_iteration_x = 32;
- break;
- }
- case DataType::QS16:
- {
- num_elems_processed_per_iteration_x = 8;
- break;
- }
-#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
- case DataType::F16:
- {
- num_elems_processed_per_iteration_x = 8;
- break;
- }
-#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
- default:
- {
- ARM_COMPUTE_ERROR("Data type not supported");
- break;
- }
- }
-
- // Configure kernel window
- Window win = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
-
- AccessWindowRectangle output_access(output->info(), 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y);
-
- update_window_and_padding(win,
- AccessWindowRectangle(input0->info(), 0, 0, 4, 1, 1.f, 0.25f),
- AccessWindowStatic(input1->info(), 0, 0, input1->info()->tensor_shape().x(), ceil_to_multiple(input1->info()->tensor_shape().y(), 4)),
- output_access);
-
- output_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), output->info()->tensor_shape()));
-
- INEKernel::configure(win);
- }
+ return Status{};
}
void NEGEMMMatrixMultiplyKernel::run(const Window &window, const ThreadInfo &info)
diff --git a/src/core/NEON/kernels/NEWeightsReshapeKernel.cpp b/src/core/NEON/kernels/NEWeightsReshapeKernel.cpp
index d52e88c..794c179 100644
--- a/src/core/NEON/kernels/NEWeightsReshapeKernel.cpp
+++ b/src/core/NEON/kernels/NEWeightsReshapeKernel.cpp
@@ -86,6 +86,57 @@
},
in);
}
+
+TensorShape get_output_shape(const ITensorInfo *input, bool has_bias)
+{
+ TensorShape output_shape{ input->tensor_shape() };
+
+ output_shape.collapse(3);
+ const size_t tmp_dim = output_shape[0];
+ output_shape.set(0, output_shape[1]);
+ output_shape.set(1, tmp_dim + (has_bias ? 1 : 0));
+
+ return output_shape;
+}
+
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *biases, const ITensorInfo *output)
+{
+ ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32);
+ ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output);
+
+ if(biases != nullptr)
+ {
+ ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases);
+ ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, biases);
+ ARM_COMPUTE_RETURN_ERROR_ON((input->num_dimensions() == 4) && (biases->num_dimensions() != 1));
+ ARM_COMPUTE_RETURN_ERROR_ON((input->num_dimensions() == 5) && (biases->num_dimensions() != 2));
+ ARM_COMPUTE_RETURN_ERROR_ON((input->num_dimensions() == 4) && (biases->dimension(0) != input->tensor_shape()[3]));
+ ARM_COMPUTE_RETURN_ERROR_ON((input->num_dimensions() == 5) && (biases->dimension(0) != input->tensor_shape()[3] || biases->dimension(1) != input->tensor_shape()[4]));
+ }
+
+ // Checks performed when output is configured
+ if(output->total_size() != 0)
+ {
+ ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), get_output_shape(input, biases != nullptr));
+ ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+ ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output);
+ }
+
+ return Status{};
+}
+
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output)
+{
+ Window window = calculate_max_window(*input, Steps());
+ window.set(Window::DimX, Window::Dimension(0, input->dimension(0), input->dimension(0)));
+ window.set(Window::DimY, Window::Dimension(0, input->dimension(1), input->dimension(1)));
+ window.set(Window::DimZ, Window::Dimension(0, input->dimension(2), input->dimension(2)));
+
+ // The NEConvolutionLayerWeightsReshapeKernel doesn't need padding so update_window_and_padding() can be skipped
+ output->set_valid_region(ValidRegion(Coordinates(), output->tensor_shape()));
+
+ return std::make_pair(Status{}, window);
+}
} // namespace
NEWeightsReshapeKernel::NEWeightsReshapeKernel()
@@ -95,35 +146,15 @@
void NEWeightsReshapeKernel::configure(const ITensor *input, const ITensor *bias, ITensor *output)
{
- ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32);
- ARM_COMPUTE_ERROR_ON_NULLPTR(output);
-
- const int fixed_point_position = input->info()->fixed_point_position();
- const DataType dt = input->info()->data_type();
- const TensorShape &input_shape = input->info()->tensor_shape();
- TensorShape output_shape{ input_shape };
- output_shape.collapse(3);
-
- const size_t tmp_dim = output_shape[0];
- output_shape.set(0, output_shape[1]);
- output_shape.set(1, tmp_dim + (bias != nullptr ? 1 : 0));
+ ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
// Output tensor auto inizialitation if not yet initialized
- auto_init_if_empty(*output->info(), output_shape, 1, dt, fixed_point_position);
+ auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(get_output_shape(input->info(), (bias != nullptr))));
- ARM_COMPUTE_ERROR_ON_MISMATCHING_DIMENSIONS(output->info()->tensor_shape(), output_shape);
- ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
- ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output);
-
- if(bias != nullptr)
- {
- ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, bias);
- ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, bias);
- ARM_COMPUTE_ERROR_ON((input->info()->num_dimensions() == 4) && (bias->info()->num_dimensions() != 1));
- ARM_COMPUTE_ERROR_ON((input->info()->num_dimensions() == 5) && (bias->info()->num_dimensions() != 2));
- ARM_COMPUTE_ERROR_ON((input->info()->num_dimensions() == 4) && (bias->info()->dimension(0) != input->info()->tensor_shape()[3]));
- ARM_COMPUTE_ERROR_ON((input->info()->num_dimensions() == 5) && (bias->info()->dimension(0) != input->info()->tensor_shape()[3] || bias->info()->dimension(1) != input->info()->tensor_shape()[4]));
- }
+ // Perform validation step
+ ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(),
+ (bias != nullptr) ? bias->info() : nullptr,
+ output->info()));
_input = input;
_bias = bias;
@@ -154,15 +185,17 @@
}
// Configure kernel
- Window window = calculate_max_window(*input->info(), Steps());
- window.set(Window::DimX, Window::Dimension(0, _input->info()->dimension(0), _input->info()->dimension(0)));
- window.set(Window::DimY, Window::Dimension(0, _input->info()->dimension(1), _input->info()->dimension(1)));
- window.set(Window::DimZ, Window::Dimension(0, _input->info()->dimension(2), _input->info()->dimension(2)));
+ auto win_config = validate_and_configure_window(input->info(), output->info());
+ ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
+ INEKernel::configure(win_config.second);
+}
- // The NEConvolutionLayerWeightsReshapeKernel doesn't need padding so update_window_and_padding() can be skipped
- output->info()->set_valid_region(ValidRegion(Coordinates(), output->info()->tensor_shape()));
+Status NEWeightsReshapeKernel::validate(const ITensorInfo *input, const ITensorInfo *biases, const ITensorInfo *output)
+{
+ ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, biases, output));
+ ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get()).first);
- INEKernel::configure(window);
+ return Status{};
}
void NEWeightsReshapeKernel::run(const Window &window, const ThreadInfo &info)