arm_compute v18.01
Change-Id: I9bfa178c2e38bfd5fc812e62aab6760d87748e05
diff --git a/src/runtime/GLES_COMPUTE/GCScheduler.cpp b/src/runtime/GLES_COMPUTE/GCScheduler.cpp
index b2235ea..fcc8559 100644
--- a/src/runtime/GLES_COMPUTE/GCScheduler.cpp
+++ b/src/runtime/GLES_COMPUTE/GCScheduler.cpp
@@ -28,11 +28,27 @@
using namespace arm_compute;
-GCScheduler::GCScheduler() = default;
+std::once_flag GCScheduler::_initialize_symbols;
+
+GCScheduler::GCScheduler()
+ : _display(EGL_NO_DISPLAY), _context(EGL_NO_CONTEXT)
+{
+}
+
+GCScheduler::~GCScheduler()
+{
+ eglDestroyContext(_display, _context);
+ eglTerminate(_display);
+
+ _context = EGL_NO_CONTEXT;
+ _display = EGL_NO_DISPLAY;
+}
void GCScheduler::default_init()
{
- GCKernelLibrary::get().init("./cs_shaders/");
+ setup_context();
+
+ GCKernelLibrary::get().init("./cs_shaders/", _display, _context);
}
void GCScheduler::init(EGLDisplay dpy, EGLContext ctx)
@@ -42,11 +58,12 @@
GCScheduler &GCScheduler::get()
{
+ std::call_once(_initialize_symbols, opengles31_is_available);
static GCScheduler scheduler;
return scheduler;
}
-void GCScheduler::enqueue(IGCKernel &kernel, bool flush)
+void GCScheduler::dispatch(IGCKernel &kernel, bool flush)
{
kernel.run(kernel.window());
if(flush)
@@ -55,7 +72,60 @@
}
}
-void GCScheduler::sync()
+void GCScheduler::memory_barrier()
{
ARM_COMPUTE_GL_CHECK(glMemoryBarrier(GL_SHADER_STORAGE_BARRIER_BIT));
}
+
+void GCScheduler::setup_context()
+{
+ EGLBoolean res;
+ _display = eglGetDisplay(EGL_DEFAULT_DISPLAY);
+
+ ARM_COMPUTE_ERROR_ON_MSG(_display == EGL_NO_DISPLAY, "Failed to get display: 0x%x.", eglGetError());
+
+ res = eglInitialize(_display, nullptr, nullptr);
+
+ ARM_COMPUTE_ERROR_ON_MSG(res == EGL_FALSE, "Failed to initialize egl: 0x%x.", eglGetError());
+ ARM_COMPUTE_UNUSED(res);
+
+ const char *egl_extension_st = eglQueryString(_display, EGL_EXTENSIONS);
+ ARM_COMPUTE_ERROR_ON_MSG((strstr(egl_extension_st, "EGL_KHR_create_context") == nullptr), "Failed to query EGL_KHR_create_context");
+ ARM_COMPUTE_ERROR_ON_MSG((strstr(egl_extension_st, "EGL_KHR_surfaceless_context") == nullptr), "Failed to query EGL_KHR_surfaceless_context");
+ ARM_COMPUTE_UNUSED(egl_extension_st);
+
+ const EGLint config_attribs[] =
+ {
+ EGL_RENDERABLE_TYPE, EGL_OPENGL_ES3_BIT_KHR,
+ EGL_NONE
+ };
+ EGLConfig cfg;
+ EGLint count;
+
+ res = eglChooseConfig(_display, config_attribs, &cfg, 1, &count);
+
+ ARM_COMPUTE_ERROR_ON_MSG(res == EGL_FALSE, "Failed to choose config: 0x%x.", eglGetError());
+ ARM_COMPUTE_UNUSED(res);
+
+ res = eglBindAPI(EGL_OPENGL_ES_API);
+
+ ARM_COMPUTE_ERROR_ON_MSG(res == EGL_FALSE, "Failed to bind api: 0x%x.", eglGetError());
+
+ const EGLint attribs[] =
+ {
+ EGL_CONTEXT_CLIENT_VERSION, 3,
+ EGL_NONE
+ };
+ _context = eglCreateContext(_display,
+ cfg,
+ EGL_NO_CONTEXT,
+ attribs);
+
+ ARM_COMPUTE_ERROR_ON_MSG(_context == EGL_NO_CONTEXT, "Failed to create context: 0x%x.", eglGetError());
+ ARM_COMPUTE_UNUSED(res);
+
+ res = eglMakeCurrent(_display, EGL_NO_SURFACE, EGL_NO_SURFACE, _context);
+
+ ARM_COMPUTE_ERROR_ON_MSG(res == EGL_FALSE, "Failed to make current: 0x%x.", eglGetError());
+ ARM_COMPUTE_UNUSED(res);
+}
diff --git a/src/runtime/GLES_COMPUTE/IGCSimpleFunction.cpp b/src/runtime/GLES_COMPUTE/IGCSimpleFunction.cpp
index 199ee46..f2926b0 100644
--- a/src/runtime/GLES_COMPUTE/IGCSimpleFunction.cpp
+++ b/src/runtime/GLES_COMPUTE/IGCSimpleFunction.cpp
@@ -38,7 +38,7 @@
{
ARM_COMPUTE_ERROR_ON_MSG(!_kernel, "The child class didn't set the GLES kernel or function isn't configured");
- GCScheduler::get().enqueue(_border_handler, false);
- GCScheduler::get().sync();
- GCScheduler::get().enqueue(*_kernel);
+ GCScheduler::get().dispatch(_border_handler, false);
+ GCScheduler::get().memory_barrier();
+ GCScheduler::get().dispatch(*_kernel);
}
diff --git a/src/runtime/GLES_COMPUTE/functions/GCArithmeticAddition.cpp b/src/runtime/GLES_COMPUTE/functions/GCArithmeticAddition.cpp
new file mode 100755
index 0000000..b99716b
--- /dev/null
+++ b/src/runtime/GLES_COMPUTE/functions/GCArithmeticAddition.cpp
@@ -0,0 +1,43 @@
+/*
+ * Copyright (c) 2016, 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/runtime/GLES_COMPUTE/functions/GCArithmeticAddition.h"
+
+#include "arm_compute/core/GLES_COMPUTE/kernels/GCArithmeticAdditionKernel.h"
+#include "support/ToolchainSupport.h"
+
+#include <utility>
+
+using namespace arm_compute;
+
+void GCArithmeticAddition::configure(const IGCTensor *input1, const IGCTensor *input2, IGCTensor *output, ConvertPolicy policy)
+{
+ auto k = arm_compute::support::cpp14::make_unique<GCArithmeticAdditionKernel>();
+ k->configure(input1, input2, output, policy);
+ _kernel = std::move(k);
+}
+
+Status GCArithmeticAddition::validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, ConvertPolicy policy)
+{
+ return GCArithmeticAdditionKernel::validate(input1, input2, output, policy);
+}
diff --git a/src/runtime/GLES_COMPUTE/functions/GCBatchNormalizationLayer.cpp b/src/runtime/GLES_COMPUTE/functions/GCBatchNormalizationLayer.cpp
index 2e546a6..99bdf43 100755
--- a/src/runtime/GLES_COMPUTE/functions/GCBatchNormalizationLayer.cpp
+++ b/src/runtime/GLES_COMPUTE/functions/GCBatchNormalizationLayer.cpp
@@ -44,5 +44,5 @@
void GCBatchNormalizationLayer::run()
{
- GCScheduler::get().enqueue(_norm_kernel, true);
+ GCScheduler::get().dispatch(_norm_kernel, true);
}
diff --git a/src/runtime/GLES_COMPUTE/functions/GCConvolutionLayer.cpp b/src/runtime/GLES_COMPUTE/functions/GCConvolutionLayer.cpp
new file mode 100644
index 0000000..5689722
--- /dev/null
+++ b/src/runtime/GLES_COMPUTE/functions/GCConvolutionLayer.cpp
@@ -0,0 +1,285 @@
+/*
+ * Copyright (c) 2017-2018 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/runtime/GLES_COMPUTE/functions/GCConvolutionLayer.h"
+
+#include "arm_compute/core/PixelValue.h"
+#include "arm_compute/core/Size2D.h"
+#include "arm_compute/core/Utils.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/runtime/GLES_COMPUTE/GCScheduler.h"
+
+#include <cmath>
+#include <memory>
+#include <tuple>
+
+using namespace arm_compute;
+
+GCConvolutionLayerReshapeWeights::GCConvolutionLayerReshapeWeights()
+ : _weights_reshape_kernel(), _weights_transposed_kernel(), _weights_reshaped(), _transpose1xW(false)
+{
+}
+
+void GCConvolutionLayerReshapeWeights::configure(const IGCTensor *weights, const IGCTensor *biases, IGCTensor *output, bool transpose1xW)
+{
+ ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::F16, DataType::F32);
+ ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(weights, output);
+ ARM_COMPUTE_ERROR_ON(weights->info()->num_dimensions() > 4);
+
+ if(biases != nullptr)
+ {
+ ARM_COMPUTE_ERROR_ON(is_data_type_quantized_asymmetric(weights->info()->data_type()));
+ ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases);
+ ARM_COMPUTE_ERROR_ON(biases->info()->dimension(0) != weights->info()->dimension(3));
+ ARM_COMPUTE_ERROR_ON(biases->info()->num_dimensions() > 1);
+ }
+
+ const bool append_biases = (biases != nullptr) && !is_data_type_quantized_asymmetric(weights->info()->data_type());
+ const unsigned bias_element = (append_biases) ? 1 : 0;
+ const IGCTensor *biases_to_use = (append_biases) ? biases : nullptr;
+
+ _transpose1xW = transpose1xW;
+
+ if(transpose1xW)
+ {
+ // Create tensor to store the reshaped weights
+ const unsigned int mat_weights_cols = weights->info()->dimension(3);
+ const unsigned int mat_weights_rows = weights->info()->dimension(0) * weights->info()->dimension(1) * weights->info()->dimension(2) + bias_element;
+ TensorShape shape_wr(mat_weights_cols, mat_weights_rows);
+ const DataType dt = weights->info()->data_type();
+ const int fixed_point_position = weights->info()->fixed_point_position();
+ TensorInfo info_wr(shape_wr, 1, dt, fixed_point_position);
+
+ _weights_reshaped.allocator()->init(info_wr);
+ _weights_reshape_kernel.configure(weights, biases_to_use, &_weights_reshaped);
+ _weights_transposed_kernel.configure(&_weights_reshaped, output);
+ _weights_reshaped.allocator()->allocate();
+ }
+ else
+ {
+ _weights_reshape_kernel.configure(weights, biases_to_use, output);
+ }
+}
+
+void GCConvolutionLayerReshapeWeights::run()
+{
+ GCScheduler::get().dispatch(_weights_reshape_kernel);
+ if(_transpose1xW)
+ {
+ GCScheduler::get().dispatch(_weights_transposed_kernel);
+ }
+}
+
+GCConvolutionLayer::GCConvolutionLayer()
+ : _reshape_weights(), _input_im2col_kernel(), _input_interleave_kernel(), _mm_kernel(), _output_col2im_kernel(), _fill_border(), _input_im2col_reshaped(), _input_interleaved_reshaped(),
+ _weights_reshaped(), _weights_transposed(), _gemm_output(), _tmp_output(), _append_bias(false), _is_fully_connected_convolution(false), _are_weights_reshaped(false)
+{
+}
+
+void GCConvolutionLayer::configure_mm(const IGCTensor *input, const IGCTensor *weights, IGCTensor *output, bool is_interleaved_transposed)
+{
+ _mm_kernel.configure(input, weights, output, 1.f, is_interleaved_transposed);
+}
+
+void GCConvolutionLayer::configure(const IGCTensor *input, const IGCTensor *weights, const IGCTensor *biases, IGCTensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info)
+{
+ ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
+ ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
+ ARM_COMPUTE_ERROR_ON(!weights_info.are_reshaped() && weights->info()->dimension(2) != input->info()->dimension(2));
+ ARM_COMPUTE_ERROR_ON(weights->info()->num_dimensions() > 4);
+
+ if(biases != nullptr)
+ {
+ ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases);
+ ARM_COMPUTE_ERROR_ON(!weights_info.are_reshaped() && biases->info()->dimension(0) != weights->info()->dimension(3));
+ ARM_COMPUTE_ERROR_ON(biases->info()->num_dimensions() > 1);
+ }
+
+ const DataType dt = input->info()->data_type();
+
+ _append_bias = (biases != nullptr);
+ _are_weights_reshaped = weights_info.are_reshaped();
+
+ const unsigned bias_element = (_append_bias) ? 1 : 0;
+ const IGCTensor *biases_to_use = (_append_bias) ? biases : nullptr;
+
+ // Get parameters from conv_info
+ unsigned int stride_x = 0;
+ unsigned int stride_y = 0;
+ std::tie(stride_x, stride_y) = conv_info.stride();
+
+ // Get convolved dimensions
+ unsigned int conv_w = 0;
+ unsigned int conv_h = 0;
+
+ const unsigned int kernel_width = (_are_weights_reshaped) ? weights_info.kernel_size().first : weights->info()->dimension(0);
+ const unsigned int kernel_height = (_are_weights_reshaped) ? weights_info.kernel_size().second : weights->info()->dimension(1);
+ std::tie(conv_w, conv_h) = scaled_dimensions(input->info()->dimension(0), input->info()->dimension(1), kernel_width, kernel_height,
+ conv_info);
+
+ // Check if its a "fully connected" convolution
+ _is_fully_connected_convolution = ((conv_w == 1) && (conv_h == 1));
+ const bool run_interleaved = (!_is_fully_connected_convolution);
+
+ unsigned int mat_weights_cols = weights->info()->dimension(3);
+ unsigned int mat_weights_rows = weights->info()->dimension(0) * weights->info()->dimension(1) * weights->info()->dimension(2) + bias_element;
+
+ // Reshape weights if needed
+ if(_are_weights_reshaped)
+ {
+ if(_is_fully_connected_convolution)
+ {
+ mat_weights_cols = weights->info()->dimension(0);
+ mat_weights_rows = weights->info()->dimension(1);
+ }
+ else
+ {
+ mat_weights_cols = weights_info.num_kernels();
+ const unsigned int quarter_reshaped_cols = weights->info()->dimension(0) / 4;
+ mat_weights_rows = quarter_reshaped_cols + bias_element;
+ }
+ }
+ else
+ {
+ if(_is_fully_connected_convolution)
+ {
+ // Create tensor to store the reshaped weights
+ int num_elems_read_per_iteration_x = 1;
+ if(dt == DataType::F16)
+ {
+ num_elems_read_per_iteration_x = 2;
+ }
+ TensorShape shape_wr((ceil_to_multiple(mat_weights_cols, num_elems_read_per_iteration_x)), mat_weights_rows);
+ _weights_reshaped.allocator()->init(weights->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_wr));
+ _reshape_weights.configure(weights, biases_to_use, &_weights_reshaped, false /* 1xW transpose */);
+ }
+ else
+ {
+ // Create tensor to store transposed weights
+ const float transpose_width = 16.0f / input->info()->element_size();
+ TensorShape shape_wt(mat_weights_rows * static_cast<unsigned int>(transpose_width), static_cast<unsigned int>(std::ceil(mat_weights_cols / transpose_width)));
+ _weights_reshaped.allocator()->init(weights->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_wt));
+ _reshape_weights.configure(weights, biases_to_use, &_weights_reshaped, true /* 1xW transpose */);
+ }
+ weights = &_weights_reshaped;
+ }
+
+ // Create tensor to store im2col reshaped inputs
+ const unsigned int mat_input_cols = mat_weights_rows;
+ const unsigned int mat_input_rows = conv_w * conv_h;
+ TensorShape shape_im2col = input->info()->tensor_shape();
+ shape_im2col.set(0, mat_input_cols);
+ shape_im2col.set(1, mat_input_rows);
+ shape_im2col.set(2, 1);
+
+ // FIXME: input->clone() doesn't work with subtensors for grouped convolutions.
+ TensorInfo im2col_reshaped_info(shape_im2col, 1, dt, input->info()->fixed_point_position());
+ _input_im2col_reshaped.allocator()->init(im2col_reshaped_info);
+
+ // Create tensor (interleave) to prepare input tensor for GEMM
+ if(run_interleaved)
+ {
+ TensorShape shape_interleaved = shape_im2col;
+ shape_interleaved.set(0, shape_interleaved.x() * 4);
+ shape_interleaved.set(1, std::ceil(shape_interleaved.y() / 4.f));
+
+ // FIXME: input->clone() doesn't work with subtensors for grouped convolutions.
+ TensorInfo interleaved_info(shape_interleaved, 1, dt, input->info()->fixed_point_position());
+ _input_interleaved_reshaped.allocator()->init(interleaved_info);
+ }
+
+ // Create GEMM output tensor
+ TensorShape shape_gemm = _input_im2col_reshaped.info()->tensor_shape();
+ shape_gemm.set(0, mat_weights_cols);
+ shape_gemm.set(1, mat_input_rows);
+ const DataType gemm_data_type = dt;
+
+ // FIXME: input->clone() doesn't work with subtensors for grouped convolutions.
+ TensorInfo info_gemm(shape_gemm, 1, gemm_data_type, input->info()->fixed_point_position());
+ _gemm_output.allocator()->init(info_gemm);
+
+ // Configure kernels
+ if(dt == DataType::F16)
+ {
+ BorderSize border_size = BorderSize(conv_info.pad_top(), conv_info.pad_right(), conv_info.pad_bottom(), conv_info.pad_left());
+ input->info()->extend_padding(border_size);
+ _fill_border.configure(input, border_size, BorderMode::CONSTANT, PixelValue(0)); // for PAD of im2col fp16: consider it as border
+ }
+ _input_im2col_kernel.configure(input, &_input_im2col_reshaped, Size2D(kernel_width, kernel_height), conv_info, _append_bias);
+
+ // Configure matrix multiply
+ if(run_interleaved)
+ {
+ _input_interleave_kernel.configure(&_input_im2col_reshaped, &_input_interleaved_reshaped);
+ configure_mm(&_input_interleaved_reshaped, weights, &_gemm_output);
+ _input_interleaved_reshaped.allocator()->allocate();
+ }
+ else
+ {
+ configure_mm(&_input_im2col_reshaped, weights, &_gemm_output, false);
+ }
+ _input_im2col_reshaped.allocator()->allocate();
+
+ // Configure Col2Im
+ _output_col2im_kernel.configure(&_gemm_output, output, std::make_pair(conv_w, conv_h));
+ _gemm_output.allocator()->allocate();
+
+ ARM_COMPUTE_ERROR_ON_MSG((output->info()->dimension(0) != conv_w) || (output->info()->dimension(1) != conv_h), "Output shape does not match the expected one");
+
+ // Allocate intermediate tensor
+ if(!_are_weights_reshaped)
+ {
+ _weights_reshaped.allocator()->allocate();
+ }
+}
+
+void GCConvolutionLayer::run()
+{
+ // Run weights reshaping (Runs once for every configure)
+ if(!_are_weights_reshaped)
+ {
+ _are_weights_reshaped = true;
+ _reshape_weights.run();
+ }
+
+ // Run im2col
+ GCScheduler::get().dispatch(_fill_border);
+ GCScheduler::get().memory_barrier();
+ GCScheduler::get().dispatch(_input_im2col_kernel);
+
+ if(!_is_fully_connected_convolution)
+ {
+ GCScheduler::get().memory_barrier();
+ // Run interleave4x4
+ GCScheduler::get().dispatch(_input_interleave_kernel);
+ }
+
+ GCScheduler::get().memory_barrier();
+ // Runs matrix multiply on reshaped matrices
+ GCScheduler::get().dispatch(_mm_kernel);
+
+ GCScheduler::get().memory_barrier();
+ // Reshape output matrix
+ GCScheduler::get().dispatch(_output_col2im_kernel, false);
+}
diff --git a/src/runtime/GLES_COMPUTE/functions/GCDepthConcatenateLayer.cpp b/src/runtime/GLES_COMPUTE/functions/GCDepthConcatenateLayer.cpp
index ee0b121..689d8be 100755
--- a/src/runtime/GLES_COMPUTE/functions/GCDepthConcatenateLayer.cpp
+++ b/src/runtime/GLES_COMPUTE/functions/GCDepthConcatenateLayer.cpp
@@ -63,7 +63,8 @@
{
for(unsigned i = 0; i < _num_inputs; i++)
{
- GCScheduler::get().enqueue(_border_handlers_vector[i], false);
- GCScheduler::get().enqueue(_concat_kernels_vector[i], true);
+ GCScheduler::get().dispatch(_border_handlers_vector[i], false);
+ GCScheduler::get().memory_barrier();
+ GCScheduler::get().dispatch(_concat_kernels_vector[i], true);
}
}
diff --git a/src/runtime/GLES_COMPUTE/functions/GCDepthwiseConvolutionLayer.cpp b/src/runtime/GLES_COMPUTE/functions/GCDepthwiseConvolutionLayer.cpp
new file mode 100644
index 0000000..ef65989
--- /dev/null
+++ b/src/runtime/GLES_COMPUTE/functions/GCDepthwiseConvolutionLayer.cpp
@@ -0,0 +1,41 @@
+/*
+ * 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/runtime/GLES_COMPUTE/functions/GCDepthwiseConvolutionLayer.h"
+
+#include "arm_compute/core/GLES_COMPUTE/IGCTensor.h"
+#include "arm_compute/core/PixelValue.h"
+#include "arm_compute/runtime/GLES_COMPUTE/GCScheduler.h"
+#include "support/ToolchainSupport.h"
+
+using namespace arm_compute;
+
+void GCDepthwiseConvolutionLayer3x3::configure(IGCTensor *input, const IGCTensor *weights, const IGCTensor *biases, IGCTensor *output, const PadStrideInfo &conv_info)
+{
+ auto k = arm_compute::support::cpp14::make_unique<GCDepthwiseConvolutionLayer3x3Kernel>();
+ k->configure(input, weights, biases, output, conv_info);
+ _kernel = std::move(k);
+
+ // Configure border handler
+ _border_handler.configure(input, _kernel->border_size(), BorderMode::CONSTANT, PixelValue(0));
+}
diff --git a/src/runtime/GLES_COMPUTE/functions/GCDropoutLayer.cpp b/src/runtime/GLES_COMPUTE/functions/GCDropoutLayer.cpp
index 032c2fd..6407464 100644
--- a/src/runtime/GLES_COMPUTE/functions/GCDropoutLayer.cpp
+++ b/src/runtime/GLES_COMPUTE/functions/GCDropoutLayer.cpp
@@ -46,5 +46,5 @@
void GCDropoutLayer::run()
{
- GCScheduler::get().enqueue(_dropout_kernel);
+ GCScheduler::get().dispatch(_dropout_kernel);
}
diff --git a/src/runtime/GLES_COMPUTE/functions/GCFullyConnectedLayer.cpp b/src/runtime/GLES_COMPUTE/functions/GCFullyConnectedLayer.cpp
index 63cb40e..9e4f0f6 100644
--- a/src/runtime/GLES_COMPUTE/functions/GCFullyConnectedLayer.cpp
+++ b/src/runtime/GLES_COMPUTE/functions/GCFullyConnectedLayer.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -61,7 +61,7 @@
_im2col_output.allocator()->init(TensorInfo(shape_im2col, 1, dt));
// Configure im2col kernel
- _im2col_kernel.configure(input, &_im2col_output, std::make_pair(1, 1), PadStrideInfo(1, 1, 0, 0), false);
+ _im2col_kernel.configure(input, &_im2col_output, Size2D(1, 1), PadStrideInfo(1, 1, 0, 0), false);
// Configure matrix multiply kernel
_mm_kernel.configure(&_im2col_output, weights, output, 1.0f, false);
@@ -159,19 +159,22 @@
// Linearize input if it comes from a convolutional layer
if(_is_fc_after_conv)
{
- GCScheduler::get().enqueue(_im2col_kernel, false);
+ GCScheduler::get().dispatch(_im2col_kernel, false);
}
- GCScheduler::get().sync();
+ if(!_are_weights_reshaped || _is_fc_after_conv)
+ {
+ GCScheduler::get().memory_barrier();
+ }
// Run matrix multiply
- GCScheduler::get().enqueue(_mm_kernel, !_accumulate_biases);
+ GCScheduler::get().dispatch(_mm_kernel, !_accumulate_biases);
// Accumulate biases if provided
if(_accumulate_biases)
{
- GCScheduler::get().sync();
+ GCScheduler::get().memory_barrier();
- GCScheduler::get().enqueue(_accumulate_biases_kernel);
+ GCScheduler::get().dispatch(_accumulate_biases_kernel);
}
}
diff --git a/src/runtime/GLES_COMPUTE/functions/GCGEMM.cpp b/src/runtime/GLES_COMPUTE/functions/GCGEMM.cpp
index c47a0e7..7aa2d42 100644
--- a/src/runtime/GLES_COMPUTE/functions/GCGEMM.cpp
+++ b/src/runtime/GLES_COMPUTE/functions/GCGEMM.cpp
@@ -38,6 +38,7 @@
#include "arm_compute/runtime/ITensorAllocator.h"
using namespace arm_compute;
+using namespace arm_compute::gles_compute;
GCGEMM::GCGEMM()
: _interleave_kernel(), _transpose_kernel(), _mm_kernel(), _ma_kernel(), _tmp_a(), _tmp_b(), _is_interleaved_transposed(false), _run_addition(false)
@@ -116,18 +117,20 @@
if(_is_interleaved_transposed)
{
// Run interleave kernel
- GCScheduler::get().enqueue(_interleave_kernel, false);
+ GCScheduler::get().dispatch(_interleave_kernel, false);
// Run transpose kernel
- GCScheduler::get().enqueue(_transpose_kernel, false);
+ GCScheduler::get().dispatch(_transpose_kernel, false);
+ GCScheduler::get().memory_barrier();
}
// Run matrix multiply kernel
- GCScheduler::get().enqueue(_mm_kernel, !_run_addition);
+ GCScheduler::get().dispatch(_mm_kernel, !_run_addition);
// Run matrix addition kernel
if(_run_addition)
{
- GCScheduler::get().enqueue(_ma_kernel);
+ GCScheduler::get().memory_barrier();
+ GCScheduler::get().dispatch(_ma_kernel);
}
}
diff --git a/src/runtime/GLES_COMPUTE/functions/GCNormalizationLayer.cpp b/src/runtime/GLES_COMPUTE/functions/GCNormalizationLayer.cpp
index d30ed52..fc3882d 100644
--- a/src/runtime/GLES_COMPUTE/functions/GCNormalizationLayer.cpp
+++ b/src/runtime/GLES_COMPUTE/functions/GCNormalizationLayer.cpp
@@ -55,7 +55,9 @@
void GCNormalizationLayer::run()
{
- GCScheduler::get().enqueue(_multiply_kernel, false);
- GCScheduler::get().enqueue(_border_handler, false);
- GCScheduler::get().enqueue(_norm_kernel, false);
+ GCScheduler::get().dispatch(_multiply_kernel, false);
+ GCScheduler::get().memory_barrier();
+ GCScheduler::get().dispatch(_border_handler, false);
+ GCScheduler::get().memory_barrier();
+ GCScheduler::get().dispatch(_norm_kernel, true);
}
diff --git a/src/runtime/GLES_COMPUTE/functions/GCNormalizePlanarYUVLayer.cpp b/src/runtime/GLES_COMPUTE/functions/GCNormalizePlanarYUVLayer.cpp
new file mode 100755
index 0000000..5fb971c
--- /dev/null
+++ b/src/runtime/GLES_COMPUTE/functions/GCNormalizePlanarYUVLayer.cpp
@@ -0,0 +1,48 @@
+/*
+ * 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/runtime/GLES_COMPUTE/functions/GCNormalizePlanarYUVLayer.h"
+
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/runtime/GLES_COMPUTE/GCScheduler.h"
+
+using namespace arm_compute;
+
+GCNormalizePlanarYUVLayer::GCNormalizePlanarYUVLayer()
+ : _norm_kernel()
+{
+}
+
+void GCNormalizePlanarYUVLayer::configure(const IGCTensor *input, IGCTensor *output, const IGCTensor *mean, const IGCTensor *sd)
+{
+ _norm_kernel.configure(input, output, mean, sd);
+}
+
+void GCNormalizePlanarYUVLayer::run()
+{
+ GCScheduler::get().dispatch(_norm_kernel, true);
+}
diff --git a/src/runtime/GLES_COMPUTE/functions/GCPoolingLayer.cpp b/src/runtime/GLES_COMPUTE/functions/GCPoolingLayer.cpp
index 46a60cd..ff03eff 100644
--- a/src/runtime/GLES_COMPUTE/functions/GCPoolingLayer.cpp
+++ b/src/runtime/GLES_COMPUTE/functions/GCPoolingLayer.cpp
@@ -23,8 +23,8 @@
*/
#include "arm_compute/runtime/GLES_COMPUTE/functions/GCPoolingLayer.h"
+#include "arm_compute/core/GLES_COMPUTE/IGCTensor.h"
#include "arm_compute/core/GLES_COMPUTE/kernels/GCPoolingLayerKernel.h"
-#include "arm_compute/core/PixelValue.h"
#include "support/ToolchainSupport.h"
using namespace arm_compute;
@@ -40,3 +40,8 @@
BorderMode border_mode = (PoolingType::MAX == pool_info.pool_type()) ? BorderMode::REPLICATE : BorderMode::CONSTANT;
_border_handler.configure(input, _kernel->border_size(), border_mode, PixelValue(0.0f));
}
+
+Status GCPoolingLayer::validate(const ITensorInfo *input, const ITensorInfo *output, const PoolingLayerInfo &pool_info)
+{
+ return GCPoolingLayerKernel::validate(input, output, pool_info);
+}
\ No newline at end of file
diff --git a/src/runtime/GLES_COMPUTE/functions/GCScale.cpp b/src/runtime/GLES_COMPUTE/functions/GCScale.cpp
new file mode 100644
index 0000000..cfe65a3
--- /dev/null
+++ b/src/runtime/GLES_COMPUTE/functions/GCScale.cpp
@@ -0,0 +1,40 @@
+/*
+ * Copyright (c) 2016, 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/runtime/GLES_COMPUTE/functions/GCScale.h"
+
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/GLES_COMPUTE/IGCTensor.h"
+#include "arm_compute/core/GLES_COMPUTE/kernels/GCScaleKernel.h"
+#include "arm_compute/core/Validate.h"
+#include "support/ToolchainSupport.h"
+
+using namespace arm_compute;
+
+void GCScale::configure(IGCTensor *input, IGCTensor *output, InterpolationPolicy policy, BorderMode border_mode, PixelValue constant_border_value, SamplingPolicy sampling_policy)
+{
+ auto k = arm_compute::support::cpp14::make_unique<GCScaleKernel>();
+ k->configure(input, output, policy, border_mode == BorderMode::UNDEFINED, sampling_policy);
+ _kernel = std::move(k);
+ _border_handler.configure(input, _kernel->border_size(), border_mode, constant_border_value);
+}
diff --git a/src/runtime/GLES_COMPUTE/functions/GCSoftmaxLayer.cpp b/src/runtime/GLES_COMPUTE/functions/GCSoftmaxLayer.cpp
index 34464ff..5221c5c 100644
--- a/src/runtime/GLES_COMPUTE/functions/GCSoftmaxLayer.cpp
+++ b/src/runtime/GLES_COMPUTE/functions/GCSoftmaxLayer.cpp
@@ -63,9 +63,9 @@
void GCSoftmaxLayer::run()
{
- GCScheduler::get().enqueue(_max_kernel, false);
- GCScheduler::get().sync();
- GCScheduler::get().enqueue(_shift_exp_sum_kernel, false);
- GCScheduler::get().sync();
- GCScheduler::get().enqueue(_norm_kernel);
+ GCScheduler::get().dispatch(_max_kernel, false);
+ GCScheduler::get().memory_barrier();
+ GCScheduler::get().dispatch(_shift_exp_sum_kernel, false);
+ GCScheduler::get().memory_barrier();
+ GCScheduler::get().dispatch(_norm_kernel);
}