COMPMID-935 - Implementing Convolution with Winograd on OpenCL (part 2)
Implemented Winograd Filter Transform 3x3 on OpenCL
Change-Id: I8f2b2dd938c5c000ef7ce392a37fb7b8b4202a4e
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/122708
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>
diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index 40aceb7..4b7fa8a 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -351,6 +351,7 @@
{ "warp_affine_bilinear", "warp_affine.cl" },
{ "warp_perspective_nearest_neighbour", "warp_perspective.cl" },
{ "warp_perspective_bilinear", "warp_perspective.cl" },
+ { "winograd_filter_transform_2x2_3x3_nchw", "winograd.cl" },
{ "winograd_input_transform_2x2_3x3_stepz1_nchw", "winograd.cl" },
{ "winograd_input_transform_2x2_3x3_stepz2_nchw", "winograd.cl" },
{ "YUYV422_to_IYUV_bt709", "color_convert.cl" },
diff --git a/src/core/CL/cl_kernels/winograd.cl b/src/core/CL/cl_kernels/winograd.cl
index fa06601..238e21a 100644
--- a/src/core/CL/cl_kernels/winograd.cl
+++ b/src/core/CL/cl_kernels/winograd.cl
@@ -205,4 +205,99 @@
vstore2(out32, 0, (__global float *)(dst_addr + 14 * dst_stride_z));
vstore2(out33, 0, (__global float *)(dst_addr + 15 * dst_stride_z));
}
-#endif //defined(NUM_TILES_X)
\ No newline at end of file
+#endif //defined(NUM_TILES_X)
+
+#if defined(NUM_CHANNELS)
+
+/** This OpenCL kernel performs Winograd filter transform 3x3 when the data format is NCHW and the output tile is 2x2
+ *
+ * @note The number of channels must be passed at compile time using -DNUM_CHANNELS: e.g. -DNUM_CHANNELS=64
+ *
+ * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32
+ * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes)
+ * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes)
+ * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes)
+ * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes)
+ * @param[in] src_step_w src_stride_w * number of elements along W processed per workitem(in bytes)
+ * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor
+ * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr
+ * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes)
+ * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes)
+ * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes)
+ * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
+ */
+__kernel void winograd_filter_transform_2x2_3x3_nchw(
+ TENSOR4D_DECLARATION(src),
+ TENSOR3D_DECLARATION(dst))
+{
+ Tensor4D src = CONVERT_TO_TENSOR4D_STRUCT(src, NUM_CHANNELS);
+
+ const __global uchar *src_addr = tensor4D_offset(&src, 0, 0, 0, 0);
+
+ // Load the values from the input tensor
+ float3 w0 = vload3(0, (__global float *)(src_addr + 0 * src_stride_y));
+ float3 w1 = vload3(0, (__global float *)(src_addr + 1 * src_stride_y));
+ float3 w2 = vload3(0, (__global float *)(src_addr + 2 * src_stride_y));
+
+ // Transform the 3x3 tile in a 4x4 tile
+ float4 out0 = 0.0f;
+ float4 out1 = 0.0f;
+ float4 out2 = 0.0f;
+ float4 out3 = 0.0f;
+
+ // Row 0
+ out0.s0 = (w0.s0);
+ out0.s1 = (w0.s0 + w0.s1 + w0.s2) * 0.5f;
+ out0.s2 = (w0.s0 + w0.s2 - w0.s1) * 0.5f;
+ out0.s3 = (w0.s2);
+
+ // Row 1
+ out1.s0 = (w0.s0 + w1.s0 + w2.s0) * 0.5f;
+ out1.s1 = (w0.s0 + w1.s0 + w2.s0 + w0.s1 + w1.s1 + w2.s1 + w0.s2 + w1.s2 + w2.s2) * 0.25f;
+ out1.s2 = (w0.s0 + w1.s0 + w2.s0 + w0.s2 + w1.s2 + w2.s2 - w0.s1 - w1.s1 - w2.s1) * 0.25f;
+ out1.s3 = (w0.s2 + w1.s2 + w2.s2) * 0.5f;
+
+ // Row 2
+ out2.s0 = (w0.s0 + w2.s0 - w1.s0) * 0.5f;
+ out2.s1 = (w0.s0 + w2.s0 + w0.s1 + w2.s1 + w0.s2 + w2.s2 - w1.s0 - w1.s1 - w1.s2) * 0.25f;
+ out2.s2 = (w0.s0 + w2.s0 + w1.s1 + w0.s2 + w2.s2 - w1.s0 - w0.s1 - w2.s1 - w1.s2) * 0.25f;
+ out2.s3 = (w0.s2 + w2.s2 - w1.s2) * 0.5f;
+
+ // Row 3
+ out3.s0 = (w2.s0);
+ out3.s1 = (w2.s0 + w2.s1 + w2.s2) * 0.5f;
+ out3.s2 = (w2.s0 + w2.s2 - w2.s1) * 0.5f;
+ out3.s3 = (w2.s2);
+
+ int z = get_global_id(2);
+ int x0 = z / NUM_CHANNELS; // idx filter
+ int y0 = z % NUM_CHANNELS; // idx channel
+
+ // Get output address
+ __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x0 * dst_stride_x + y0 * dst_stride_y;
+
+ // Store the 16 values across the 16 channels
+ *(__global float *)(dst_addr + 0 * dst_stride_z) = out0.s0;
+ *(__global float *)(dst_addr + 1 * dst_stride_z) = out0.s1;
+ *(__global float *)(dst_addr + 2 * dst_stride_z) = out0.s2;
+ *(__global float *)(dst_addr + 3 * dst_stride_z) = out0.s3;
+ *(__global float *)(dst_addr + 4 * dst_stride_z) = out1.s0;
+ *(__global float *)(dst_addr + 5 * dst_stride_z) = out1.s1;
+ *(__global float *)(dst_addr + 6 * dst_stride_z) = out1.s2;
+ *(__global float *)(dst_addr + 7 * dst_stride_z) = out1.s3;
+ *(__global float *)(dst_addr + 8 * dst_stride_z) = out2.s0;
+ *(__global float *)(dst_addr + 9 * dst_stride_z) = out2.s1;
+ *(__global float *)(dst_addr + 10 * dst_stride_z) = out2.s2;
+ *(__global float *)(dst_addr + 11 * dst_stride_z) = out2.s3;
+ *(__global float *)(dst_addr + 12 * dst_stride_z) = out3.s0;
+ *(__global float *)(dst_addr + 13 * dst_stride_z) = out3.s1;
+ *(__global float *)(dst_addr + 14 * dst_stride_z) = out3.s2;
+ *(__global float *)(dst_addr + 15 * dst_stride_z) = out3.s3;
+}
+#endif // defined(NUM_CHANNELS)
diff --git a/src/core/CL/kernels/CLWinogradFilterTransformKernel.cpp b/src/core/CL/kernels/CLWinogradFilterTransformKernel.cpp
new file mode 100644
index 0000000..3dbbe15
--- /dev/null
+++ b/src/core/CL/kernels/CLWinogradFilterTransformKernel.cpp
@@ -0,0 +1,139 @@
+/*
+ * Copyright (c) 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/core/CL/kernels/CLWinogradFilterTransformKernel.h"
+
+#include "arm_compute/core/AccessWindowStatic.h"
+#include "arm_compute/core/CL/CLHelpers.h"
+#include "arm_compute/core/CL/CLHelpers.h"
+#include "arm_compute/core/CL/CLKernelLibrary.h"
+#include "arm_compute/core/CL/ICLTensor.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/IAccessWindow.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/Utils.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/Window.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+
+#include "support/ToolchainSupport.h"
+
+using namespace arm_compute;
+using namespace arm_compute::misc::shape_calculator;
+
+namespace
+{
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output)
+{
+ ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
+ ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(0) != 3);
+ ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(0) != input->dimension(1));
+ ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() > 4);
+
+ // Checks performed when output is configured
+ if(output->total_size() != 0)
+ {
+ const TensorInfo tensor_info_output = input->clone()->set_tensor_shape(compute_winograd_filter_transform_shape(*input));
+
+ ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &tensor_info_output);
+ ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+ }
+
+ return Status{};
+}
+
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output)
+{
+ ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+
+ constexpr unsigned int num_elems_processed_per_iteration_x = 3;
+ constexpr unsigned int num_elems_processed_per_iteration_y = 3;
+
+ Window win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
+ bool window_changed = false;
+
+ AccessWindowRectangle input_access(input, 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y);
+ AccessWindowStatic output_access(output, 0, 0, output->dimension(0), output->dimension(1));
+ window_changed = update_window_and_padding(win, input_access, output_access);
+ output_access.set_valid_region(win, input->valid_region());
+
+ Window win_collapsed = win.collapse(win, Window::DimZ);
+
+ Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
+ return std::make_pair(err, win_collapsed);
+}
+} // namespace
+
+CLWinogradFilterTransformKernel::CLWinogradFilterTransformKernel()
+ : _input(nullptr), _output(nullptr)
+{
+}
+
+void CLWinogradFilterTransformKernel::configure(const ICLTensor *input, ICLTensor *output)
+{
+ ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+
+ // Output tensor auto inizialitation if not yet initialized
+ auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(compute_winograd_filter_transform_shape(*input->info())));
+
+ ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info()));
+
+ // Set build options
+ CLBuildOptions build_opts;
+ build_opts.add_option("-DNUM_CHANNELS=" + support::cpp11::to_string(input->info()->dimension(2)));
+
+ // Create kernel
+ _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("winograd_filter_transform_2x2_3x3_nchw", build_opts.options()));
+
+ _input = input;
+ _output = output;
+
+ // Configure kernel window
+ auto win_config = validate_and_configure_window(input->info(), output->info());
+ ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
+ ICLKernel::configure(win_config.second);
+}
+
+Status CLWinogradFilterTransformKernel::validate(const ITensorInfo *input, const ITensorInfo *output)
+{
+ ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output));
+ ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get()).first);
+
+ return Status{};
+}
+
+void CLWinogradFilterTransformKernel::run(const Window &window, cl::CommandQueue &queue)
+{
+ ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+ ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
+
+ // Setup output window
+ Window window_out;
+ window_out.use_tensor_dimensions(_output->info()->tensor_shape(), 0);
+
+ unsigned int idx = 0;
+ add_4D_tensor_argument(idx, _input, window);
+ add_3D_tensor_argument(idx, _output, window_out);
+ enqueue(queue, *this, window);
+}