COMPMID-1277 - Optimizing CLIm2ColKernel for NHWC.
This patch includes:
- Im2Col optimizations for NHWC using a new data layout
- Refactoring of CLIm2ColKernel adding validation method and auto-init
- Removed im2col_reduced from CLIm2ColKernel and created a new kernel CLFlattenLayerKernel
Change-Id: I1620640b6796baa268324b33ae92cdd8de53e27c
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/141241
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>
diff --git a/arm_compute/core/CL/CLKernels.h b/arm_compute/core/CL/CLKernels.h
index 737d8df..ea942bd 100644
--- a/arm_compute/core/CL/CLKernels.h
+++ b/arm_compute/core/CL/CLKernels.h
@@ -62,6 +62,7 @@
#include "arm_compute/core/CL/kernels/CLErodeKernel.h"
#include "arm_compute/core/CL/kernels/CLFastCornersKernel.h"
#include "arm_compute/core/CL/kernels/CLFillBorderKernel.h"
+#include "arm_compute/core/CL/kernels/CLFlattenLayerKernel.h"
#include "arm_compute/core/CL/kernels/CLFloorKernel.h"
#include "arm_compute/core/CL/kernels/CLGEMMInterleave4x4Kernel.h"
#include "arm_compute/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.h"
diff --git a/arm_compute/core/CL/kernels/CLFlattenLayerKernel.h b/arm_compute/core/CL/kernels/CLFlattenLayerKernel.h
new file mode 100644
index 0000000..a5cf6e0
--- /dev/null
+++ b/arm_compute/core/CL/kernels/CLFlattenLayerKernel.h
@@ -0,0 +1,74 @@
+/*
+ * 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.
+ */
+#ifndef __ARM_COMPUTE_CLFLATTENLAYERKERNEL_H__
+#define __ARM_COMPUTE_CLFLATTENLAYERKERNEL_H__
+
+#include "arm_compute/core/CL/ICLKernel.h"
+
+namespace arm_compute
+{
+class ICLTensor;
+
+/** OpenCL interface for the flatten kernel.*/
+class CLFlattenLayerKernel : public ICLKernel
+{
+public:
+ /** Default constructor */
+ CLFlattenLayerKernel();
+ /** Prevent instances of this class from being copied (As this class contains pointers) */
+ CLFlattenLayerKernel(const CLFlattenLayerKernel &) = delete;
+ /** Prevent instances of this class from being copied (As this class contains pointers) */
+ CLFlattenLayerKernel &operator=(const CLFlattenLayerKernel &) = delete;
+ /** Allow instances of this class to be moved */
+ CLFlattenLayerKernel(CLFlattenLayerKernel &&) = default;
+ /** Allow instances of this class to be moved */
+ CLFlattenLayerKernel &operator=(CLFlattenLayerKernel &&) = default;
+ /** Set the input and output of the kernel.
+ *
+ * @param[in] input First input tensor to flatten with at least 3 dimensions.
+ * The dimensions above the third will be interpreted as batches. Data types supported: U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32
+ * @param[out] output Output tensor with shape [w*h*d, input_batches] where:
+ * w = width input tensor, h = height input tensor and d = depth input tensor. Data type supported: same as @p input
+ */
+ void configure(const ICLTensor *input, ICLTensor *output);
+ /** Static function to check if given info will lead to a valid configuration of @ref CLFlattenLayerKernel
+ *
+ * @param[in] input First input tensor to flatten with at least 3 dimensions.
+ * The dimensions above the third will be interpreted as batches. Data types supported: U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32
+ * @param[out] output Output tensor with shape [w*h*d, input_batches] where:
+ * w = width input tensor, h = height input tensor and d = depth input tensor. Data type supported: same as @p input
+ *
+ * @return a status
+ */
+ static Status validate(const ITensorInfo *input, const ITensorInfo *output);
+
+ // Inherited methods overridden:
+ void run(const Window &window, cl::CommandQueue &queue) override;
+
+public:
+ const ICLTensor *_input;
+ ICLTensor *_output;
+};
+} // namespace arm_compute
+#endif /*__ARM_COMPUTE_CLFLATTENLAYERKERNEL_H__ */
diff --git a/arm_compute/core/CL/kernels/CLIm2ColKernel.h b/arm_compute/core/CL/kernels/CLIm2ColKernel.h
index fc930ab..ae19319 100644
--- a/arm_compute/core/CL/kernels/CLIm2ColKernel.h
+++ b/arm_compute/core/CL/kernels/CLIm2ColKernel.h
@@ -96,48 +96,13 @@
// Inherited methods overridden:
void run(const Window &window, cl::CommandQueue &queue) override;
-private:
- /** Run the reshape kernel optimised for the special case (stride is 1, padding is 0 and kernel's low 3 dimensions are same as input)
- *
- * @param[in] window Region on which to execute the kernel. (Must be a valid region of the window returned by window()).
- * @param[in,out] queue Command queue on which to enqueue the kernel.
- */
- void run_reduced(const Window &window, cl::CommandQueue &queue);
- /** run the generic convolution layer input reshape kernel
- *
- * @param[in] window Region on which to execute the kernel. (Must be a valid region of the window returned by window()).
- * @param[in,out] queue Command queue on which to enqueue the kernel.
- */
- void run_generic(const Window &window, cl::CommandQueue &queue);
-
- /** Chooses and configure the right kernel for the given input arguments.
- *
- * @param[in] input The input tensor to convert. 3 lower dimensions represent a single input [width, height, IFM],
- * while every optional dimension from 4 and above represent a batch of inputs. Data types supported: QASYMM8/F16/F32
- * @param[in] output The output tensor. First 2 lower dimensions represent a transform of each 3D input,
- * while every dimension above represents a batch. Data types supported: Same as @p input
- * @param[in] kernel_dims The kernel dimensions (width and height).
- * @param[in] dilation Dilation, in elements, across x and y. Defaults to (1, 1).
- * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo.
- * @param[in] has_bias In case biases are provided expands the matrix with 1.
- * @param[out] build_opts OpenCL buil program options.
- *
- * @return the name of the kernel chosen
- */
- std::string configure_window(const ICLTensor *input, ICLTensor *output, const Size2D &kernel_dims,
- const Size2D &dilation, const PadStrideInfo &conv_info, CLBuildOptions &build_opts);
-
- /** Common signature for the kernel to run */
- using Im2ColFunction = void (CLIm2ColKernel::*)(const Window &, cl::CommandQueue &);
-
public:
const ICLTensor *_input;
ICLTensor *_output;
- PadStrideInfo _conv_info;
std::pair<unsigned int, unsigned int> _convolved_dims;
- unsigned int _num_elems_processed_per_iteration;
- Im2ColFunction _run_func;
- Size2D _kernel_dims;
+ unsigned int _num_elems_processed_per_iteration;
+ Size2D _kernel_dims;
+ PadStrideInfo _conv_info;
};
} // namespace arm_compute
#endif /*__ARM_COMPUTE_CLIM2COLKERNEL_H__ */
diff --git a/arm_compute/core/utils/misc/ShapeCalculator.h b/arm_compute/core/utils/misc/ShapeCalculator.h
index 1e5b9af..0a2a535 100644
--- a/arm_compute/core/utils/misc/ShapeCalculator.h
+++ b/arm_compute/core/utils/misc/ShapeCalculator.h
@@ -215,19 +215,13 @@
return output_shape;
}
-inline TensorShape compute_im2col_fc_shape(const ITensorInfo *input, const int num_input_dimensions = 3)
+inline TensorShape compute_flatten_shape(const ITensorInfo *input)
{
+ // The output shape will be the flatten version of the input (i.e. [ width * height * channels, num_batches, ... ] ). Used for FlattenLayer and FullyConnectedLayer.
+
TensorShape output_shape{ input->tensor_shape() };
- output_shape.collapse(num_input_dimensions);
-
- return output_shape;
-}
-inline TensorShape compute_im2col_flatten_shape(const ITensorInfo *input)
-{
- // The output shape will be the flatten version of the input (i.e. [ width * height * channels, 1, 1, ... ] ). Used for FlattenLayer.
- TensorShape output_shape{ input->tensor_shape() };
- output_shape.collapse(3, 0);
+ output_shape.collapse(3);
return output_shape;
}
diff --git a/arm_compute/runtime/CL/functions/CLFlattenLayer.h b/arm_compute/runtime/CL/functions/CLFlattenLayer.h
index 88df4a7..ebc0e5e 100644
--- a/arm_compute/runtime/CL/functions/CLFlattenLayer.h
+++ b/arm_compute/runtime/CL/functions/CLFlattenLayer.h
@@ -33,7 +33,7 @@
/** Basic function to execute flatten. This function calls the following OpenCL kernel:
*
-* -# @ref CLIm2ColKernel
+* -# @ref CLFlattenLayerKernel
*
*/
class CLFlattenLayer : public ICLSimpleFunction
@@ -41,11 +41,22 @@
public:
/** Initialise the kernel's input and output.
*
- * @param[in] input First input tensor to flatten with at least 3 dimensions. The dimensions over the third will be interpreted as batches. Data types supported: F16/F32
+ * @param[in] input First input tensor to flatten with at least 3 dimensions.
+ * The dimensions above the third will be interpreted as batches. Data types supported: U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32
* @param[out] output Output tensor with shape [w*h*d, input_batches] where:
- * w = width input tensor, h = height input tensor and d = depth input tensor. Data type supported: same as @p input
+ * w = width input tensor, h = height input tensor and d = depth input tensor. Data type supported: same as @p input
*/
void configure(const ICLTensor *input, ICLTensor *output);
+ /** Static function to check if given info will lead to a valid configuration of @ref CLTranspose
+ *
+ * @param[in] input First input tensor to flatten with at least 3 dimensions.
+ * The dimensions above the third will be interpreted as batches. Data types supported: U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32
+ * @param[out] output Output tensor with shape [w*h*d, input_batches] where:
+ * w = width input tensor, h = height input tensor and d = depth input tensor. Data type supported: same as @p input
+ *
+ * @return a status
+ */
+ static Status validate(const ITensorInfo *input, const ITensorInfo *output);
};
} // namespace arm_compute
diff --git a/arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h b/arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h
index e8fe8e4..450cd83 100644
--- a/arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h
+++ b/arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h
@@ -27,11 +27,11 @@
#include "arm_compute/runtime/CL/ICLSimpleFunction.h"
#include "arm_compute/core/CL/kernels/CLGEMMMatrixAccumulateBiasesKernel.h"
-#include "arm_compute/core/CL/kernels/CLIm2ColKernel.h"
#include "arm_compute/core/CL/kernels/CLTransposeKernel.h"
#include "arm_compute/runtime/CL/CLMemoryGroup.h"
#include "arm_compute/runtime/CL/CLTensor.h"
#include "arm_compute/runtime/CL/functions/CLConvertFullyConnectedWeights.h"
+#include "arm_compute/runtime/CL/functions/CLFlattenLayer.h"
#include "arm_compute/runtime/CL/functions/CLGEMM.h"
#include "arm_compute/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.h"
#include "arm_compute/runtime/CL/functions/CLGEMMLowpOutputStage.h"
@@ -130,14 +130,14 @@
void configure_mm(const ICLTensor *input, const ICLTensor *weights, ICLTensor *output);
CLMemoryGroup _memory_group;
- CLIm2ColKernel _im2col_kernel;
CLConvertFullyConnectedWeights _convert_weights;
+ CLFlattenLayer _flatten_layer;
CLFullyConnectedLayerReshapeWeights _reshape_weights_kernel;
CLGEMM _mm_gemm;
CLGEMMLowpMatrixMultiplyCore _mm_gemmlowp;
CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint _gemmlowp_output_stage;
CLGEMMMatrixAccumulateBiasesKernel _accumulate_biases_kernel;
- CLTensor _im2col_output;
+ CLTensor _flatten_output;
CLTensor _gemmlowp_output;
CLTensor _converted_weights_output;
CLTensor _reshape_weights_output;