COMPMID-631: Merge branches/gles_compute branch

Last commit:
commit b25c5f68042b0c81bf611d59a1bb8535e1c42497
Author: Xinghang Zhou <xinghang.zhou@arm.com>
Date:   Wed Oct 25 18:48:10 2017 +0800

    Synced validation's tolerances of GCSoftmax from cl side

Change-Id: Ibe72054205c1c8721845d679a31af7ed0a7c5cf6
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/93283
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>

1 files changed, 128 insertions, 0 deletions

diff --git a/src/core/GLES_COMPUTE/kernels/GCGEMMTranspose1xWKernel.cpp b/src/core/GLES_COMPUTE/kernels/GCGEMMTranspose1xWKernel.cpp
new file mode 100644
index 0000000000..a1270b4c3d
--- /dev/null
+++ b/src/core/GLES_COMPUTE/kernels/GCGEMMTranspose1xWKernel.cpp
@@ -0,0 +1,128 @@
+/*
+ * Copyright (c) 2017 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/core/GLES_COMPUTE/kernels/GCGEMMTranspose1xWKernel.h"
+
+#include "arm_compute/core/AccessWindowTranspose.h"
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/GLES_COMPUTE/GCHelpers.h"
+#include "arm_compute/core/GLES_COMPUTE/GCKernelLibrary.h"
+#include "arm_compute/core/GLES_COMPUTE/IGCTensor.h"
+#include "arm_compute/core/GLES_COMPUTE/OpenGLES.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/Window.h"
+
+#include <cmath>
+
+using namespace arm_compute;
+
+void GCGEMMTranspose1xWKernel::configure(const IGCTensor *input, IGCTensor *output)
+{
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_NULLPTR(output);
+
+    TensorShape  output_shape{ input->info()->tensor_shape() };
+    const size_t transpose_w = 16 / input->info()->element_size();
+    output_shape.set(0, input->info()->dimension(1) * transpose_w);
+    output_shape.set(1, static_cast<size_t>(std::ceil((input->info()->dimension(0) / static_cast<float>(transpose_w)))));
+
+    // Output tensor auto inizialitation if not yet initialized
+    auto_init_if_empty(*output->info(), output_shape, 1, input->info()->data_type(), input->info()->fixed_point_position());
+
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_DIMENSIONS(output->info()->tensor_shape(), output_shape);
+
+    const unsigned int num_elems_processed_per_iteration = 16 / input->info()->element_size();
+    const int          scale_x                           = num_elems_processed_per_iteration;
+
+    _input  = input;
+    _output = output;
+
+    std::set<std::string> build_opts;
+    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(1));
+    build_opts.emplace("#define LOCAL_SIZE_Y " + support::cpp11::to_string(1));
+    build_opts.emplace("#define LOCAL_SIZE_Z " + support::cpp11::to_string(1));
+    /*
+     * Following an example of how the transposition1xW works when the input data type is F32
+     *
+     *         |a00 a01 a02 a03|
+     *         |a10 a11 a12 a13|
+     *         |a20 a21 a22 a23| = | a00 a01 a02 a03 || a10 a11 a12 a13 || a20 a21 a22 a23 || a30 a31 a32 a33 |
+     *         |a30 a31 a32 a33|
+     *
+     * The output matrix will have the following shape: [ height * W, ceil(width / W) ], where W = (16 / element size of the tensor)
+     */
+    // Create kernel
+    build_opts.emplace("#define GEMM_TRANSPOSE1xW");
+    _kernel = GCKernelLibrary::get().create_kernel("gemm_transpose1x4", build_opts);
+
+    // Configure window
+    Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration));
+
+    ARM_COMPUTE_ERROR_ON_MSG((win.x().end() / scale_x) == 0, "Transposed shape would be 0 in the second dimension");
+
+    AccessWindowHorizontal input_access(input->info(), 0, num_elems_processed_per_iteration);
+    AccessWindowTranspose  output_access(output->info(), 0, 0, num_elems_processed_per_iteration, 1, scale_x, 1.f / scale_x);
+
+    update_window_and_padding(win, input_access, output_access);
+
+    output_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), input->info()->tensor_shape()));
+
+    _kernel.clear_params();
+    // set shader params binding point
+    _kernel.set_shader_params_binding_point(0);
+
+    IGCKernel::configure(win);
+}
+
+void GCGEMMTranspose1xWKernel::run(const Window &window)
+{
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IGCKernel::window(), window);
+
+    _kernel.use();
+
+    // Output is transposed
+    Window out_window(window);
+    out_window.set(Window::DimX, window.y());
+    out_window.set(Window::DimY, window.x());
+
+    Window in_slice  = window.first_slice_window_2D();
+    Window out_slice = out_window.first_slice_window_2D();
+
+    do
+    {
+        unsigned int idx = 0;
+        add_2D_tensor_argument(idx, _input, 1, in_slice);
+        add_2D_tensor_argument(idx, _output, 2, out_slice);
+
+        _kernel.update_shader_params();
+
+        enqueue(*this, in_slice);
+    }
+    while(window.slide_window_slice_2D(in_slice) && out_window.slide_window_slice_2D(out_slice));
+}