COMPMID-617: Add validate support for NEON FullyConnectedLayer

Change-Id: I08987022c8d4cc335c00b8af27bd3edb8fe64d3b
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/111596
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Alexander Gilday <alexander.gilday@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

3 files changed, 154 insertions, 40 deletions

diff --git a/src/core/NEON/kernels/NEGEMMMatrixAccumulateBiasesKernel.cpp b/src/core/NEON/kernels/NEGEMMMatrixAccumulateBiasesKernel.cpp
index 3dd59bddd6..cab3c7a58f 100644
--- a/src/core/NEON/kernels/NEGEMMMatrixAccumulateBiasesKernel.cpp
+++ b/src/core/NEON/kernels/NEGEMMMatrixAccumulateBiasesKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -39,6 +39,42 @@
 
 using namespace arm_compute;
 
+namespace
+{
+inline Status validate_arguments(const ITensorInfo *accum, const ITensorInfo *biases)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(accum, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(biases, accum);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT_POSITION(biases, accum);
+    ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1);
+    ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != accum->dimension(0));
+
+    return Status{};
+}
+
+inline std::pair<Status, Window> validate_and_configure_window(ITensorInfo *accum, ITensorInfo *biases)
+{
+    constexpr unsigned int num_elems_processed_per_iteration = 16;
+
+    // Configure kernel window
+    Window win = calculate_max_window(*accum, Steps(num_elems_processed_per_iteration));
+
+    bool window_changed = update_window_and_padding(win,
+                                                    AccessWindowHorizontal(accum, 0, num_elems_processed_per_iteration),
+                                                    AccessWindowStatic(biases, 0, 0, ceil_to_multiple(biases->dimension(0), num_elems_processed_per_iteration), biases->tensor_shape().y()));
+
+    AccessWindowHorizontal output_access(accum, 0, num_elems_processed_per_iteration);
+
+    // Set the valid region for the accum tensor
+    Coordinates coord;
+    coord.set_num_dimensions(accum->num_dimensions());
+    output_access.set_valid_region(win, ValidRegion(coord, accum->tensor_shape()));
+
+    Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
+    return std::make_pair(err, win);
+}
+} // namespace
+
 NEGEMMMatrixAccumulateBiasesKernel::NEGEMMMatrixAccumulateBiasesKernel()
     : _accum(nullptr), _biases(nullptr)
 {
@@ -46,31 +82,26 @@ NEGEMMMatrixAccumulateBiasesKernel::NEGEMMMatrixAccumulateBiasesKernel()
 
 void NEGEMMMatrixAccumulateBiasesKernel::configure(ITensor *accum, const ITensor *biases)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(accum, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32);
-    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(biases, accum);
-    ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT_POSITION(biases, accum);
-    ARM_COMPUTE_ERROR_ON(biases->info()->num_dimensions() > 1);
+    ARM_COMPUTE_ERROR_ON_NULLPTR(accum, biases);
+
+    // Perform validate step
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(accum->info(), biases->info()));
 
     _biases = biases;
     _accum  = accum;
 
-    constexpr unsigned int num_elems_processed_per_iteration = 16;
-
     // Configure kernel window
-    Window win = calculate_max_window(*accum->info(), Steps(num_elems_processed_per_iteration));
-
-    update_window_and_padding(win,
-                              AccessWindowHorizontal(accum->info(), 0, num_elems_processed_per_iteration),
-                              AccessWindowStatic(biases->info(), 0, 0, ceil_to_multiple(biases->info()->dimension(0), num_elems_processed_per_iteration), biases->info()->tensor_shape().y()));
-
-    AccessWindowHorizontal output_access(accum->info(), 0, num_elems_processed_per_iteration);
+    auto win_config = validate_and_configure_window(accum->info(), biases->info());
+    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
+    INEKernel::configure(win_config.second);
+}
 
-    // Set the valid region for the accum tensor
-    Coordinates coord;
-    coord.set_num_dimensions(accum->info()->num_dimensions());
-    output_access.set_valid_region(win, ValidRegion(coord, accum->info()->tensor_shape()));
+Status NEGEMMMatrixAccumulateBiasesKernel::validate(const ITensorInfo *accum, const ITensorInfo *biases)
+{
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(accum, biases));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(accum->clone().get(), biases->clone().get()).first);
 
-    INEKernel::configure(win);
+    return Status{};
 }
 
 void NEGEMMMatrixAccumulateBiasesKernel::run(const Window &window, const ThreadInfo &info)
diff --git a/src/core/NEON/kernels/NEGEMMMatrixMultiplyKernel.cpp b/src/core/NEON/kernels/NEGEMMMatrixMultiplyKernel.cpp
index aa5e2dd0dd..69b052a9bd 100644
--- a/src/core/NEON/kernels/NEGEMMMatrixMultiplyKernel.cpp
+++ b/src/core/NEON/kernels/NEGEMMMatrixMultiplyKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -36,6 +36,8 @@
 #include "arm_compute/core/Validate.h"
 #include "arm_compute/core/Window.h"
 
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+
 #include <arm_neon.h>
 #include <cstddef>
 #include <cstdint>
@@ -1409,27 +1411,73 @@ void matrix_matrix_multiply_qs16(const ITensor *input0, const ITensor *input1, I
     ina, inb, out);
 }
 
-Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output)
+inline Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output, float alpha, bool is_interleaved, const GEMMReshapeInfo &reshape_info)
 {
+    ARM_COMPUTE_UNUSED(alpha);
+
     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)
+    if(!is_interleaved)
     {
         ARM_COMPUTE_RETURN_ERROR_ON(input0->dimension(0) != input1->dimension(1));
+
+        if(output->total_size() != 0)
+        {
+            ARM_COMPUTE_RETURN_ERROR_ON(input1->dimension(0) != output->dimension(0));
+            ARM_COMPUTE_RETURN_ERROR_ON(input0->dimension(1) != output->dimension(1));
+            ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, output);
+            ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input0, output);
+        }
+    }
+    else
+    {
+        const int m                         = reshape_info.m();
+        const int n                         = reshape_info.n();
+        const int k                         = reshape_info.k();
+        const int mult_transpose1xW_width   = reshape_info.mult_transpose1xW_width();
+        const int mult_interleave4x4_height = reshape_info.mult_interleave4x4_height();
+
+        /* Interleave */
+        TensorShape tensor_shape0{ input0->tensor_shape() };
+        tensor_shape0.set(0, k);
+        tensor_shape0.set(1, m);
+
+        const TensorInfo tensor_info0          = input0->clone()->set_tensor_shape(tensor_shape0);
+        const TensorInfo tensor_info_reshaped0 = input0->clone()->set_tensor_shape(misc::shape_calculator::compute_interleaved_shape(tensor_info0, mult_interleave4x4_height));
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input0, &tensor_info_reshaped0);
+
+        if(n != 0) /* Transpose */
+        {
+            TensorShape tensor_shape1{ input1->tensor_shape() };
+            tensor_shape1.set(0, n);
+            tensor_shape1.set(1, k);
+
+            const TensorInfo tensor_info1          = input1->clone()->set_tensor_shape(tensor_shape1);
+            const TensorInfo tensor_info_reshaped1 = input1->clone()->set_tensor_shape(misc::shape_calculator::compute_transpose1xW_with_element_size_shape(tensor_info1, mult_transpose1xW_width));
+            ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input1, &tensor_info_reshaped1);
+        }
+
+        if(output->total_size() != 0)
+        {
+            if(n != 0)
+            {
+                ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(0) != static_cast<size_t>(n));
+            }
+            ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(1) != static_cast<size_t>(m));
+            ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, output);
+            ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input0, output);
+        }
     }
 
     return Status{};
 }
 
-std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input0, ITensorInfo *input1, ITensorInfo *output)
+inline std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input0, ITensorInfo *input1, ITensorInfo *output)
 {
-    Window win            = Window();
-    bool   window_changed = false;
+    bool   window_changed{};
+    Window win{};
 
     unsigned int       num_elems_processed_per_iteration_x = 0;
     const unsigned int num_elems_processed_per_iteration_y = 4;
@@ -1538,11 +1586,19 @@ NEGEMMMatrixMultiplyKernel::NEGEMMMatrixMultiplyKernel()
 {
 }
 
-void NEGEMMMatrixMultiplyKernel::configure(const ITensor *input0, const ITensor *input1, ITensor *output, float alpha)
+void NEGEMMMatrixMultiplyKernel::configure(const ITensor *input0, const ITensor *input1, ITensor *output, float alpha, bool is_interleaved, const GEMMReshapeInfo &reshape_info)
 {
-    // Perform validate step
     ARM_COMPUTE_ERROR_ON_NULLPTR(input0, input1, output);
-    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input0->info(), input1->info(), output->info()));
+
+    // Output tensor auto inizialitation if not yet initialized
+    TensorShape tensor_shape{ input0->info()->tensor_shape() };
+    tensor_shape.set(0, is_interleaved ? reshape_info.n() : input1->info()->dimension(0));
+    tensor_shape.set(1, is_interleaved ? reshape_info.m() : input0->info()->dimension(1));
+
+    auto_init_if_empty(*output->info(), input0->info()->clone()->set_tensor_shape(tensor_shape));
+
+    // Perform validate step
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input0->info(), input1->info(), output->info(), alpha, is_interleaved, reshape_info));
 
     _input0 = input0;
     _input1 = input1;
@@ -1555,9 +1611,10 @@ void NEGEMMMatrixMultiplyKernel::configure(const ITensor *input0, const ITensor
     INEKernel::configure(win_config.second);
 }
 
-Status NEGEMMMatrixMultiplyKernel::validate(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output)
+Status NEGEMMMatrixMultiplyKernel::validate(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output, float alpha, bool is_interleaved,
+                                            const GEMMReshapeInfo &reshape_info)
 {
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input0, input1, output));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input0, input1, output, alpha, is_interleaved, reshape_info));
     ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input0->clone().get(), input1->clone().get(), output->clone().get()).first);
 
     return Status{};
diff --git a/src/core/NEON/kernels/NEIm2ColKernel.cpp b/src/core/NEON/kernels/NEIm2ColKernel.cpp
index 633f78de4b..4fa329bf44 100644
--- a/src/core/NEON/kernels/NEIm2ColKernel.cpp
+++ b/src/core/NEON/kernels/NEIm2ColKernel.cpp
@@ -32,6 +32,8 @@
 #include "arm_compute/core/Types.h"
 #include "arm_compute/core/Validate.h"
 
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+
 #include <arm_neon.h>
 #include <cstddef>
 #include <cstdint>
@@ -42,14 +44,34 @@ using namespace arm_compute;
 
 namespace
 {
-Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias)
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info,
+                          bool has_bias, bool is_fully_connected, bool is_flatten)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QASYMM8, DataType::QS16, DataType::F16, DataType::F32);
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT_POSITION(input, output);
     ARM_COMPUTE_RETURN_ERROR_ON(input->data_type() == DataType::QASYMM8 && has_bias);
-    ARM_COMPUTE_UNUSED(kernel_dims);
-    ARM_COMPUTE_UNUSED(conv_info);
+
+    if(is_flatten) /* Called by FlattenLayer */
+    {
+        size_t flatten_shape = input->tensor_shape().x() * input->tensor_shape().y() * input->tensor_shape().z();
+        ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(0) != flatten_shape);
+    }
+    else if(!is_fully_connected) /* Called by ConvolutionLayer */
+    {
+        std::pair<unsigned int, unsigned int> out_dims = scaled_dimensions(input->dimension(0), input->dimension(1), kernel_dims.width, kernel_dims.height, conv_info);
+        ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(0) != (input->dimension(2) * kernel_dims.area() + (has_bias ? 1 : 0)));
+        ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(1) != (out_dims.first * out_dims.second));
+        ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(2) != 1);
+    }
+    else /* Called by FullyConnectedLayer */
+    {
+        const int num_batch_dimensions = std::max(0, static_cast<int>(output->tensor_shape().num_dimensions()) - 1);
+        const int num_input_dimensions = input->tensor_shape().num_dimensions() - num_batch_dimensions;
+
+        TensorInfo expected_output = output->clone()->set_tensor_shape(misc::shape_calculator::compute_im2col_shape(input, num_input_dimensions));
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(&expected_output, output);
+    }
 
     return Status{};
 }
@@ -291,12 +313,15 @@ NEIm2ColKernel::NEIm2ColKernel()
 {
 }
 
-void NEIm2ColKernel::configure(const ITensor *input, ITensor *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias)
+void NEIm2ColKernel::configure(const ITensor *input, ITensor *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info,
+                               bool has_bias, bool is_fully_connected, bool is_flatten)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
 
     // Perform validation step
-    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), kernel_dims, conv_info, has_bias));
+    ARM_COMPUTE_UNUSED(is_fully_connected);
+    ARM_COMPUTE_UNUSED(is_flatten);
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), kernel_dims, conv_info, has_bias, is_fully_connected, is_flatten));
 
     _input          = input;
     _output         = output;
@@ -382,9 +407,10 @@ void NEIm2ColKernel::configure(const ITensor *input, ITensor *output, const Size
     IKernel::configure(window);
 }
 
-Status NEIm2ColKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias)
+Status NEIm2ColKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info,
+                                bool has_bias, bool is_fully_connected, bool is_flatten)
 {
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, kernel_dims, conv_info, has_bias));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, kernel_dims, conv_info, has_bias, is_fully_connected, is_flatten));
     return Status{};
 }