APPBROWSER-377: GCConvoutionLayer support for FP16

Change-Id: I801b5e393a16a9f92c062826e6fcfd5982ca7bb3
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/116584
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

1 files changed, 47 insertions, 22 deletions

diff --git a/src/core/GLES_COMPUTE/kernels/GCCol2ImKernel.cpp b/src/core/GLES_COMPUTE/kernels/GCCol2ImKernel.cpp
index 492f708a98..af1e34ef59 100644
--- a/src/core/GLES_COMPUTE/kernels/GCCol2ImKernel.cpp
+++ b/src/core/GLES_COMPUTE/kernels/GCCol2ImKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -24,6 +24,7 @@
 
 #include "arm_compute/core/GLES_COMPUTE/kernels/GCCol2ImKernel.h"
 
+#include "arm_compute/core/AccessWindowStatic.h"
 #include "arm_compute/core/Error.h"
 #include "arm_compute/core/GLES_COMPUTE/GCHelpers.h"
 #include "arm_compute/core/GLES_COMPUTE/GCKernelLibrary.h"
@@ -43,31 +44,50 @@ GCCol2ImKernel::GCCol2ImKernel()
 void GCCol2ImKernel::configure(const IGCTensor *input, IGCTensor    *output,
                                std::pair<unsigned int, unsigned int> convolved_dims)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_NULLPTR(output);
+
+    TensorShape output_shape = input->info()->tensor_shape();
+    output_shape.set(0, convolved_dims.first);
+    output_shape.set(1, convolved_dims.second);
+    output_shape.set(2, input->info()->tensor_shape()[0]);
+
+    // Output auto inizialitation if not yet initialized
+    auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(output_shape));
+
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_DIMENSIONS(output->info()->tensor_shape(), output_shape);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
 
     _input          = input;
     _output         = output;
     _convolved_dims = convolved_dims;
 
+    unsigned int num_elems_processed_per_iteration = 1;
+
     // Create kernel
-    std::set<std::string>  build_opts;
-    constexpr unsigned int num_elems_processed_per_iteration = 8;
+    std::set<std::string> build_opts;
+    build_opts.emplace("#define WIDTH_OUTPUT " + support::cpp11::to_string(_convolved_dims.first));
+    std::string dt_name = (input->info()->data_type() == DataType::F32) ? "DATA_TYPE_FP32" : "DATA_TYPE_FP16";
+    build_opts.emplace(("#define " + dt_name));
     build_opts.emplace("#define LOCAL_SIZE_X " + support::cpp11::to_string(num_elems_processed_per_iteration));
     build_opts.emplace("#define LOCAL_SIZE_Y " + support::cpp11::to_string(num_elems_processed_per_iteration));
-    build_opts.emplace("#define LOCAL_SIZE_Z " + support::cpp11::to_string(1));
-    build_opts.insert("#define COL2IM");
-    _kernel = static_cast<GCKernel>(GCKernelLibrary::get().create_kernel("col2im", build_opts));
+    build_opts.emplace("#define LOCAL_SIZE_Z " + support::cpp11::to_string(num_elems_processed_per_iteration));
 
-    // Set static kernel arguments
-    unsigned int idx = num_arguments_per_2D_tensor() + num_arguments_per_3D_tensor();
-    _kernel.set_argument(idx++, _convolved_dims.first);
+    _kernel = static_cast<GCKernel>(GCKernelLibrary::get().create_kernel("col2im", build_opts));
 
     // Configure window
-    Window win = calculate_max_window(*input->info(), Steps());
+    unsigned int nums = 2;
+    Window       win  = calculate_max_window(*output->info(), Steps(nums));
 
-    // The GCCol2ImKernel doesn't need padding so update_window_and_padding() can be skipped
-    output->info()->set_valid_region(ValidRegion(Coordinates(), output->info()->tensor_shape()));
+    AccessWindowHorizontal output_access(output->info(), 0, 2);
+    const int              input_padding = ceil_to_multiple(input->info()->dimension(0), 2) - input->info()->dimension(0);
+
+    AccessWindowStatic input_access(input->info(), 0, 0, input->info()->dimension(0) + input_padding, input->info()->dimension(1) + 1);
+
+    update_window_and_padding(win, input_access,
+                              output_access);
+
+    output_access.set_valid_region(win, output->info()->valid_region());
 
     IGCKernel::configure(win);
 }
@@ -77,20 +97,25 @@ void GCCol2ImKernel::run(const Window &window)
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_WINDOWS(IGCKernel::window(), window);
 
-    Window slice_in  = window.first_slice_window_2D();
-    Window slice_out = window.first_slice_window_3D();
-
     _kernel.use();
 
+    Window collapsed_window = window.collapse_if_possible(IGCKernel::window(), Window::DimZ);
+    Window slice            = collapsed_window.first_slice_window_3D();
+
+    // Set static kernel arguments
+    unsigned int idx = 2 * num_arguments_per_3D_tensor();
+    //_kernel.set_argument(idx++, _output->info()->strides_in_bytes()[3]);
+    _kernel.set_argument(idx++, uint(_output->info()->dimension(2)));
+    _kernel.set_argument(idx++, _input->info()->strides_in_bytes()[2]);
+
     do
     {
         // Set inputs
-        unsigned int idx     = 0;
-        unsigned int binding = 1;
-        add_2D_tensor_argument(idx, _input, binding++, slice_in);
-        add_3D_tensor_argument(idx, _output, binding++, slice_out);
+        unsigned int idx = 0;
+        add_2D_tensor_argument(idx, _input, 1, slice);
+        add_3D_tensor_argument(idx, _output, 2, slice);
         _kernel.update_shader_params();
-        enqueue(*this, slice_in);
+        enqueue(*this, slice);
     }
-    while(window.slide_window_slice_2D(slice_in) && window.slide_window_slice_3D(slice_out));
+    while(collapsed_window.slide_window_slice_3D(slice));
 }